Repository navigation
Release v0.4.0 - #466
Merged
Merged
Release v0.4.0#466
Conversation
A request with no Credential now carries at most 64 KB, read apart from everyone else's in a pool of 64, and a Secret naming nobody is answered before its body is read. Kamosu runs its own accept loop so hyper has a header deadline (20 s); a whole request has ten minutes, a stranger's body 30 s, and past 256 requests at once Kamosu answers busy. An Operation sent without a JSON Content-Type is refused, and so is a request from a page on another address. The Session cookie becomes __Host-kamosu_session, and every answer gains Permissions-Policy, COOP, CORP, and HSTS when sent to an https public address. The offline outbox keeps a write that meets a busy answer rather than dropping it. The numbers and their reasons live in src/admission.rs. Closes #368 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01UC2Kj954nT9iSHXhT8qJPX
Close the front door to strangers on a bare port (Closes #368)
The agent beside a Runner reads pages strangers wrote, so a page that tricks it could have it send away the provider's key it held. The Runner now holds the key, given as KAMOSU_RUNNER_MODEL_KEY with the one address it may go to, and opens a relay for each Job's agent alone: a loopback port for an agent command, a socket for an agent in a container of its own. The agent sends its model requests there with a stand-in key, and the relay puts the real key where the stand-in was (Authorization, x-api-key or x-goog-api-key), sends the request under that address only and streams the answer back. It closes when the Job ends. The published agent image points opencode at the relay, and can be started as a fence whose network the agent shares: the internet and Kamosu's address, never the host or the LAN. An agent that still holds its own key keeps working, and both logs say to move it. Closes #372 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01UC2Kj954nT9iSHXhT8qJPX
The Runner holds the model's key and relays the agent's requests (closes #372)
An Operator's Access Key used to carry every Operator power, so any agent handed one could rewrite the House Notes every later Job reads. A key now acts as Operator only when an Operator signed in mints it with `operator: true`, and loses that for good when its Person steps down. Migration 54 makes every live key of a current Operator an Operator key, so nothing set up before the upgrade stops working. Also: a read-only key fetches no Backup; a recovery link ends every Session and Access Key of the account; `delete_kitchen` asks the Credential rather than the account. Settings offers an Operator the choice (option C with option B's warning), and a form that sets a password asks for it twice. Closes #369 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01UC2Kj954nT9iSHXhT8qJPX
An Access Key acts as Operator only when minted as an Operator key (Closes #369)
A person can log in with a passkey beside the password: the device offers the passkeys it holds for the public address, with no name typed. Settings lists each passkey above the devices, adds one after the password, renames and removes them. A passkey works only at the public address's host, and only where that is an https:// name; the Operator's address setting says how many a new address would strand. A person can also change their own password, choosing whether their other devices are signed out and whether their Access Keys end. webauthn-rs 0.6 is held at its prerelease: 0.5 links OpenSSL. The test phone is written by hand on p256, since the passkey client library turns on serde_json's preserve_order for the whole build. ADR 0051 records the rest. Closes #370 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01UC2Kj954nT9iSHXhT8qJPX
Log in with a passkey, and change your own password (closes #370)
- A start as another user first clears any Runner, agent and fence still running in the developer's own Podman, which would ask for the same Jobs with the developer's keys in reach. - The caps allow no swap: --memory-swap equals --memory, where Podman would otherwise grant the same amount again as swap. - runner-logs takes its log followers down with it on Ctrl-C. - A settings file with no agent lines no longer stops runner-start without a message. - _runner-as is now _check-runner-as, the comments are shorter, and AGENTS.md lists the variable. The operator side applies the decision "Each public self-built app runs as its own unprivileged owner" (ADR 0032). Closes #381 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01KMCiGZSCLtWC32obHepGxw
feat: the dev Runner recipes can run as another local user (#381)
) A recipe PDF is read by Kamosu started again as `kamosu read-pdf`, held to 512 MB of memory, 15 s of processor time and no core dump, and stopped at 10 s; an instance reads two at once and answers busy past that. Read in place, a 218 KB crafted PDF grew the dev server to 10 GB. A Crouton export unpacks at most 4 GB across its files, damaged ones included, and the Import Report names each recipe past that as not read. Decoding a photograph, making a Display Copy and drawing a share card run off the request workers. A count of wrong passwords is forgotten an hour after its last miss, at every name alike (migration 56, ADR 0031 amended), so made-up names no longer pile up in login_failures and nothing tells real names apart. Closes #373 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01F6Lk6ZvL7sewZuHZoe1c9y
Bound what a PDF, a Crouton export or a guessed name can cost (closes #373)
runner-start points the Runner, and the agent through the fence, at the dev instance on the host's LAN address. Podman 5 gives a rootless container pasta by default, and pasta copies the host's address into the container, so that address reaches the container itself and the connection is refused. Naming slirp4netns keeps the path Podman 4.9 takes today. Checked on Podman 4.9 as kamosu-dev: the three start, the Runner asks the dev instance, and from the agent the dev port answers 200, another LAN host is refused and the internet answers 200, as before. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01E8UHAt45LTz72u6qPNQFKM
fix: the dev Runner and its fence name slirp4netns as their network
…riting them Ubuntu 26.04's uutils install cannot write /dev/stdin over a file that already exists, so every run after the first failed with "install: No such file or directory". The recipe rewrites both files on every run, so removing them first loses nothing. Closes #387 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01JQ6ZTjv2e4C5z3g5wK76uR
…install fix: runner-start as another user survives uutils install (Closes #387)
…s a build that died The tokenizer pads every input in a call to the longest one, and a recipe's Note is one line, whole, so one long Note padded every short line of its recipe to its own length. A recipe of 40 short lines and a 1,758-character Note added 390 MB during indexing, enough for a capped container to be killed on every start. The embedder now counts each input's tokens with the model's own tokenizer and sends them shortest first, in groups that hold no more padded tokens than one input of the maximum length (2,048), measured to add at most ~50 MB. The vectors come back in the caller's order. A build notes which item it hands the model in a new column, meaning_search.embedding_now (migration 57), and clears it once its rows are written. A build that finds one left behind logs a warning naming it before embedding anything. Closes #389 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01JQ6ZTjv2e4C5z3g5wK76uR
fix: Meaning Search embeds a recipe in groups of one length, and names a build that died
Closes #371. Each Person reads their own record in Settings › Security record: each sign-in and how it was made, a run of wrong passwords at their name as one row, passkeys, password changes, Access Keys and Sessions ended. Every Operator also reads the instance's: the House Notes and the public address changed, who administers, accounts disabled or deleted, Operator keys, Runner Keys, Backups fetched, and a crowd of sign-ins refused at once, each naming who did it through which Session or key. Both are Operations, so an agent reads them too. Kept 90 days, swept daily. Every event is also one `kamosu-security` line on standard output, in the fixed format docs/security-record.md describes. It names Credentials, never an address, a password or a Secret, and Kamosu sends it nowhere. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01JQ6ZTjv2e4C5z3g5wK76uR
feat: a Security Record of what happens at the doors (Closes #371)
A workflow runs `just audit` on each pull request into dev and every Monday against dev. A weekly failure opens one issue, or comments on it while it stays open. The token that can write issues lives in a job of its own, apart from the code the audit installs. npm audit has nowhere to record an advisory judged not to apply, so ui/npm-audit-exceptions.json does what deny.toml does for Rust, each entry with its reason, and ui/tools/npm-audit.mjs judges npm's report against it. Its first entry is the braces advisory published on 2026-09-18. It has no patched version and reaches Kamosu only through Tailwind's file watcher, which Kamosu never starts. Closes #374 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01JQ6ZTjv2e4C5z3g5wK76uR
The first live run opened its issue, but a search straight after did not return it: GitHub's search index lags a new issue. The report now lists the open issues and matches the title exactly. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01JQ6ZTjv2e4C5z3g5wK76uR
just audit runs on every pull request into dev and weekly (closes #374)
A Runner Key now records the Person who minted it. Setting a new password with a recovery link ends every live Runner Key that account minted and, for an Operator, every one minted before Kamosu recorded the maker. Each is revoked as by hand: its held Job waits again and its Job Keys stop working. Job Keys are otherwise left alone, since they go to a Runner and never to the person who asked. The recovery's answer and its signed_in entry count ended_runner_keys, every Operator's record notes each key as runner_key_revoked by the recovered Person, and the recovery page warns an Operator beforehand. Closes #378 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01126KLKw9pGwH2vG3r1i2VL
…-and-job-keys A recovery link ends the Runner Keys minted with the old password (Closes #378)
Signing out was one button on this device's row of the Sessions list, three pages down, and the owner could not find it. It now stands at the foot of the Settings menu, under who is signed in, on both layouts (option A of 20 October 2026). Both ways in run one sheet. It first sends what the device wrote offline, which signing out would otherwise lose without a word. If some cannot go, it says how many and offers to wait and try again before signing out anyway. Otherwise it asks, saying the recipes kept for offline use leave the device. Offline both buttons wait for the server. A Session already ended elsewhere goes straight to the sign-in screen. Prior art is in docs/research/signing-out-in-other-apps.md. Closes #380 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01126KLKw9pGwH2vG3r1i2VL
Sign out at the foot of Settings, asking first (Closes #380)
In an English line, a comma with a digit before it and exactly three digits after it separates thousands, so `1,020 g` is stored and read as 1020 g. `1,5 g` and `1,0200` keep their decimal comma, and French and Spanish lines are read as before (#180). `read_line` takes the Branch's Language to know which. Reading the library again reads each Version in the Language it was set to last, and a Translation's source in its own Language. Closes #383 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01126KLKw9pGwH2vG3r1i2VL
Import a KitchenOwl export (closes #417)
An administrator's Mealie backup is read from its tables into the shape of Mealie's export and through the export's own reader, so both give the same recipes under the same ids. Each line is written as Mealie writes it, from the quantity, unit, food and note. A backup of several households asks which is the cook's, in a sheet (option B), and keeps the upload for the answer. Each "I Made This" becomes a finished Attempt of the importer, beginning "Cooked by <name> in Mealie", recorded in imported_attempts (migration 61) so none arrives twice. No password or token column is read. Every era of Mealie's anonymised samples is read by its revision. The Mealie half of #434 is done too: a total time, a time Kamosu cannot read and nutrition beyond calories go into the Note, for the export as for the backup. ADR 0053 amends ADR 0005 and ADR 0010. Closes #418 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01126KLKw9pGwH2vG3r1i2VL
Import a Mealie backup, with its cooking history as Attempts (Closes #418)
ctutils 0.4.3 -> 0.4.5 (RUSTSEC-2026-0385), tokio 1.53.1 -> 1.54.1 (RUSTSEC-2026-0386) and the test-only crypto-bigint 0.7.5 -> 0.7.6 (RUSTSEC-2026-0387), each by `cargo update -p`. Unlocking tokio also has Cargo re-pick windows-sys 0.52.0, already in the tree, for six crates that accept either; that touches Windows builds alone. crypto-bigint 0.5.5 keeps RUSTSEC-2026-0387 under webauthn-rs, which no 0.5 release fixes, and is excused in deny.toml: passkey sign-in only checks signatures against public keys. undici 8.11.2 -> 8.12.3 under jsdom, a test dependency, by `npm update undici` (GHSA-xjj6-f33q-929c). The braces excuse goes: GitHub withdrew GHSA-vfj7-8cjw-p6xm on 2026-10-23. No changelog line: passkeys are still unreleased, and v0.3.0 held neither ctutils nor webauthn-rs. Closes #432 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01126KLKw9pGwH2vG3r1i2VL
Clear the dependency audit (Closes #432)
Paprika's total time, a time Kamosu reads only in part or not at all, servings no Yield says whole, the difficulty and the free-text nutrition each go into the Note as a line of their own, in the recipe's Language. Crouton's nutrition text stays out of the recipe (ADR 0025) and is now counted on the Import Report. Closes #434 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01126KLKw9pGwH2vG3r1i2VL
Keep what Paprika holds and no field does in the Note (closes #434)
ADR 0054: an iPhone Shortcut Kamosu publishes, Android's share target, one add_this Operation that decides everything, and an Add-only key the Shortcut carries. The glossary gains the Add-only key and an Original of several pictures; ADR 0031 points to the new key kind. The prior-art survey behind it goes under docs/research/. Refs #452 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
ADR 0031, 0044 and 0048 take amendment notes rather than edits. ADR 0054 says which ADRs it amends and writes the Add-only key as the glossary does. It records the session's defaults, refuses the key on the routes outside the Catalogue, and waits up to ten seconds on a link so its message can say what happened. The prior-art claims now match their sources. Refs #452 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
search_recipes takes `cooking`: never cooked, cooked with no rating, or a rating word as some Person's latest rating, any of them held. It reads everyone the reader cooks with, the company the recipe page counts, and the reader alone with `mine`, which then judges the whole shelf rather than cutting it to their history first. On the Recipes screen every filter goes behind one row of buttons under the search box, Cookbooks, Kitchens, Cooking and Tags, each naming what it holds and opening its list, as chosen on the ticket. ADR 0027 is amended to say so. Closes #445 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01126KLKw9pGwH2vG3r1i2VL
Filter the shelf by how a recipe was cooked (Closes #445)
docs: record how a phone shares a recipe into Kamosu
A Step links the line of the section it works in, the food itself before what is made from it, and a word of the title when the line is that food. A word for what the recipe makes never links a line alone, and neither does a food named for what something is for or in a when clause. A Step naming a section's ingredients links every line of it. set_step_uses corrects which lines one Step uses, kept on later Versions while the Step's words and those lines are unchanged (ADR 0054). Closes #439 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01126KLKw9pGwH2vG3r1i2VL
fix: link the line a Step adds, and let a Step's lines be corrected (#439)
An iPhone shares through an Apple Shortcut that carries an Add-only key, an Access Key that may call add_this and get_job and stage uploads, and nothing else. Android shares into Kamosu itself once installed, through a share target the service worker answers. Both hand everything to one new Operation, add_this, which decides where a link, text, photos or a PDF go and words its answer in the asker's Reading Language. Several photos now go to the agent as the pages of one recipe: an Original is an ordered set of files, and reading again and sending again use all of them. Settings gains Share from your phone, which mints the key and shows its setup code once, and the key list marks an Add-only key and changes where it sends what it shares. The decision record is renumbered 0055: 0054 was taken by #439. Closes #452 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01126KLKw9pGwH2vG3r1i2VL
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01126KLKw9pGwH2vG3r1i2VL
Add a recipe from a phone's Share menu (closes #452)
…k it Building the Shortcut from docs/sharing-from-an-iphone.md and sharing to it from a phone found four places where the steps fail: - Get Type names a Safari page "Safari Web Page", not "Safari web page", and the comparison minds capitals, so a Safari page went up as a file. The names are also translated. Step 8 now tests that the name contains "Safari" and takes the page's link from Get Details of Safari Web Page. - Get URLs from Input finds nothing in a Safari page, and a URL is its own link, so step 7 takes the input itself. - A key path such as result.upload_id that misses stops the Shortcut with Apple's "Could not evaluate the key path", so a refused upload hid Kamosu's message. Every answer is now read one key at a time, and a refused upload shows Kamosu's message and stops. - Converting every image to JPEG blackens a PNG's clear parts. Only a HEIC or HEIF is converted now. Every case in "Checking it before publishing" passed on the dev instance with the Shortcut built this way, which is now published: its iCloud link is SHORTCUT_URL and the runbook's last line. The shared copy carries no setup code. Refs #452 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01NabmVcHX5k7RZnnd21CZiK
docs: make the Shortcut runbook match the Shortcut as tested, and link it
A Photograph and an Original are encoded as WebP without transparency, and encode_at dropped the alpha channel with to_rgb8, which leaves each clear pixel the colour it stores. Most PNGs store black there, so a recipe card of dark text on a clear background became a black square, and the agent read nothing on it. A picture with an alpha channel is now laid on white first. A picture without one is encoded exactly as before, so no existing Photograph's bytes, and so none of its deduplication, change. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01NabmVcHX5k7RZnnd21CZiK
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01NabmVcHX5k7RZnnd21CZiK
…n-black fix: lay a picture's clear parts on white, not black
… hand A source line whose Reading a person set with an amount is now compared as it was set, so `2 heads broccoli` corrected to two broccoli with no unit agrees with `2 têtes de brocoli` and the Translation takes *broccoli*. Any other source line is read again, as before, so a correction that named only the Food keeps crossing over. Closes #444 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01126KLKw9pGwH2vG3r1i2VL
A Translation's line is measured against the source's Reading set by hand (#444)
Each Step on the edit screen gets an Ingredients button opening a sheet of tick boxes over every Ingredient Line (option B of 25 October 2026). The ticks go to set_step_uses after the save, on the Version it left, by where lines and Steps sit in what was saved. A Step whose words or corrected lines change in the same edit is read again instead. get_recipe answers as_read on a corrected Step, which is what putting the correction back returns to. Closes #456 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01126KLKw9pGwH2vG3r1i2VL
…redients Correct which ingredients a Step uses from the edit screen (closes #456)
Sets the version to 0.4.0 in Cargo.toml, Cargo.lock, ui/package.json, ui/package-lock.json, and dates the changelog's Unreleased section as v0.4.0. Made by `just release-prepare 0.4.0`.
…h v0.4.0 The release checklist found the README and Kamosu's Docker Hub page still naming 0.3.0 and missing imports from other apps, sharing from a phone, passkeys, the Security Record and the HTTP/1.1 rule. The glossary lacked the Shortcut, the Operator's new powers, and the Runner's relay and fence; ADR 0052 now records that a Demo copies nothing to a phone. The changelog cites #405, and the agent image's rebuild line moves to 2026-10-25 for Alpine's musl, ca-certificates and brotli fixes. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01126KLKw9pGwH2vG3r1i2VL
Prepare the v0.4.0 release
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Releases v0.4.0: merges
devintomain.The release was reviewed on the version-bump pull request, #465.
just release-publish 0.4.0merges this pull request, and refuses while any box in that pull request's checklist is unticked. Publishing waits for the maintainer to approve the release title and notes.