Yioop — Development process for LLM co-implementers

How Chris Pollett likes development to run, written for an LLM working with him on Yioop. Most of this is shared with his other projects; what is specific to Yioop — the gold/work sync, the createdb check, the locale and library rules — is marked. When something says "always" or "never," treat it as a hard rule.

Round counter. Keep a count of rounds since you last re-read this file. A round is one exchange: Chris says something, you respond, and that includes every time he stops you mid-work, however short what he says. Echo the current count in bold as the first thing in every response. When the count reaches 10, re-read this whole file before writing your next response. Write 2-3 sentences for each section in this file briefly describing it and how you can adhere to that section more closely in the next round. Then reset the count to 0. A re-read Chris asks for before the count reaches 10 never resets it: the count goes on, and the review at 10 is still done. This is a memory aid: it keeps these conventions fresh over a long session instead of letting them drift. The same counter runs in atto's copy of this file.

A section's description says what the section requires: its commands, files, rules and numbers, so that someone who has never opened this file learns what it asks. A one-line maxim is not a description. “Calls that free what frees itself are never written” told Chris nothing about which calls those are or why; the functions, the PHP version that made them do nothing, and the script that finds them are the description. Most of this file is meant to serve Chris's other projects too, so describe the general rule first and the Yioop-specific part as Yioop's, and leave out details that change over time, such as which site runs which server.

The review goes in the body of the response where Chris reads it, never in a file he has to open. Writing it to a file and saying it exists is not writing it.

Two things go in the review beyond the section-by-section reading, and both are the point of the exercise rather than an addition to it:

Every lapse in the ten rounds, named, with what stops it recurring. Not a general resolution to do better: the rule that was broken, the round it was broken in, and the mechanical guard or changed habit that catches it next time. A lapse Chris had to find and report is worth naming first.
What the ten rounds show about the process itself. Where rounds went to waste, which check caught a fault late that could have caught it early, which step cost more than it returned. Say what to change in this file, and change it in the same patch.

How changes move (Yioop-specific): gold and work

Yioop is developed against a two-clone discipline:

gold is the confirmed line of history. Its confirmed-head tag marks the latest confirmed commit.
work is a throwaway clone of gold where changes are made. It is re-cloned from gold after every gold update, with the configured work_directory restored into it.

Setting up gold, and keeping it current, from seekquarry.com. The confirmed history lives at https://www.seekquarry.com/group/2/yioop-repo.git, which now serves smart-HTTP clones, so gold is a shallow clone of it rather than a hand-built local vault. To create gold, shallow-clone to a couple hundred commits of history: git clone --depth 200 https://www.seekquarry.com/group/2/yioop-repo.git into the gold path. That is enough recent history for context without pulling the whole repository. Tag confirmed-head at the clone's tip.

Each round you deliver one patch. Gold then advances not by pulling from the remote each round — that has proven unreliable over the shallow clone — but by applying your own delivered patch to gold once Chris confirms he has committed it. The signal is his showing you the commit message that goes with your patch: if his reply names or quotes that commit message, the patch is confirmed, so apply it to gold — git apply the delivery, commit it in gold with that message, and move confirmed-head to the new commit — then re-clone work. If his reply shows no commit message for it, the patch is not confirmed, so leave gold exactly as it was and do not apply it. This keeps gold's tree in step with what Chris committed without depending on a network pull; the local commit hash will differ from his, but the tree content matches, which is all a later confirmed-head..HEAD diff needs.

gold is a vault, not a workbench. Exactly two operations ever touch gold: cloning it to make work, and advancing it by applying a patch Chris has confirmed and moving confirmed-head to it, as above. Nothing else, ever. Never run a build, a test, a golden regeneration, a ConfigureTool command such as createdb, or any other command inside gold; never create a work_directory there; never hand-edit or hand-copy a file into it, and the only commit it ever takes is that one confirmed patch. Every build, test, and regeneration happens in a work clone. If you ever find yourself about to go into gold to run something — even just to check a result or to get back in sync — that urge is the process breaking; stop, and do it in work.

The loop each round, and this must be done prior to any new work: first settle gold. If Chris's reply shows the commit message for the patch you delivered last round, apply that patch to gold, move confirmed-head to the new commit, re-clone work, and restore work_directory; if his reply shows no such commit message, gold and work stay as they are and you build this round's work on top of what is already there. Starting new work without settling gold this way is the most common way the two lines drift. If a round's work has already begun and only then does Chris confirm the previous patch, do not re-clone work — that would discard the in-progress changes; instead apply the confirmed patch to gold, move confirmed-head to that commit in both gold and work, and only then cut. Cut a delivery as git diff confirmed-head..HEAD, and dry-run it with git apply --check on a fresh clone checked out at confirmed-head before presenting. Apply patches forward only — never reverse-diff. Never assume anything is committed until Chris shows the commit message for it. Name a patch for what it does in plain terms, never a sequence number or a plan-item number.

Two scripts that keep gold in step with what Chris committed

Saying gold advanced is not advancing it. The failure this guards against is writing “gold is now at X” without having run the apply, which leaves gold behind Chris's tree; the next patch is then cut against the wrong base and fails on his machine. Two small scripts in devlog/ remove the judgment from the reporting, and a ledger file records what has been delivered and what has been settled.

devlog/devscripts/settle_gold.sh "commit message" is the first action of every round, always run, with the commit message Chris showed or with no argument when he showed none. With a message it applies the oldest delivered-but-unsettled patch to gold, commits it, moves confirmed-head, and marks it settled. Either way it prints gold's head read back from the tree, and any patch still waiting. The gold line in your response is that printed output, pasted. If you did not run it, you have nothing to paste, which is the whole point. Run it in a call of its own and read its whole output: a filter such as grep on that output once hid the gold line, and a round then went out with no gold line to paste.
settle_gold.sh --withdraw name-of-patch takes a delivered patch Chris did not apply out of the ledger. It is that round's one settle call: it goes on to count the round and print gold's head, as a settle with no message does.
A delivered patch Chris neither confirms with a commit message nor rejects is folded into the next patch. That round's settle is --withdraw with the patch's name, and the round's patch carries both changes under a new name. Waiting for the commit message instead held a fix back two rounds.
devlog/devscripts/cut_patch.sh name-of-patch cuts the delivery and refuses while anything in the ledger is still waiting. That refusal is what makes a stale gold impossible to build on: a dry-run against a fresh clone of gold cannot catch a stale gold, because the clone is gold and passes happily.

The ledger lives outside the repository (by default /home/claude/gold_ledger.txt) and holds one line per patch: its path, its name, and delivered or settled. Both scripts take GOLD, WORK, OUT and LEDGER from the environment, so a sandbox laid out differently only needs those set.

A bug that outlives a conversation

A conversation can reach its limit mid-chase. When one does, write the state of the chase as a file and hand it over — what is established, what is ruled out, the plan and which step it is on, the exact next action, and the traps met along the way. Such a note stays out of the repository: it is scaffolding for one hunt, not something the history should carry forever, so it goes to Chris as a file of its own and never into a patch. A new session reads it after this process file and carries on without repeating the ruled-out ground.

Setting up the dev environment (Yioop-specific)

Beyond the gold/work clones above, a work clone needs a configured work_directory before it will build or run. Let php be the command you use for the most recent version of PHP you have access to (needs to be > PHP 8.1). Restore the configured work_directory into the fresh work clone, point Yioop at it with php src/configs/ConfigureTool.php work-dir, then build the two SQLite databases with php src/configs/ConfigureTool.php createdb, which should print “Create DB succeeded.” Without a configured work_directory, Config.php stops short of defining its later constants (for example MAX_GROUP_PAGE_LEN) and createdb dies with an undefined-constant error — that failure is the environment, not your patch. Strings and resources are always read from src/locale first. work_directory/app/locale (LOCALE_DIR) may be absent, or hold only the strings and resources a site changed, which take the place of the shipped ones. Nothing writes to src/locale: TokenTool and the Manage Locales screen write to app/locale.

Tools that ship with Yioop

src/executables/CodeTool.php is the general-purpose tool. Run a test script with php src/executables/CodeTool.php unit <TestFile> (or unit list / unit ClassName methodName to target cases); check docblocks with needsdocs <file>; and check the 80-column rule with longlines <file> (multibyte-aware, so it is the authority over a raw byte count).
Yioop itself can be run under its own web server via the index.php script in the root directory of the project, or under another web server such as nginx or Apache.
There are many Command-line tools that can easily be installed in this environment: lynx and curl for fetching and reading pages as text, plus node and npm. Rather than say don't have the tool you want, try to install it.
For a fault in an index or in search results, run php src/executables/QueryTool.php '<query> i:<timestamp>' explain and php src/executables/ArcTool.php dict <bundle> <word> before writing any script of your own. The plan shows which key a query reads and how many documents it holds, and the dictionary shows each partition's rows. Seven rounds once went to a google query a home-made dictionary reader could not explain; these two commands explained it in one.
For a slow page or file, measure before changing anything. The excimer sampling profiler builds here from its source with phpize and loads with php -d extension=...; one run named the two methods a thumbnail spent nine seconds in. A change to how a thumbnail is drawn is checked by drawing the thumbnails of every test document before and after and comparing them pixel by pixel.

Getting a headless browser

wkhtmltoimage and wkhtmltopdf are installed and turn a page or fragment into a PNG or PDF with no X server, but they are an old WebKit: they skip CSS nesting and resolve var(), clamp(), and some @media interplay unlike a current browser, so a width they report can be a renderer artifact. Use them for a quick look at a fragment; when a measurement has to be right, use a real browser.

The rules for a walk hold for every server a round starts, whether it serves work or a copy of Yioop set up in a scratch folder for one test: a kill timeout at launch, errors shown, pages read with lynx -dump, the cache off, settings put back and stop_test_servers.sh at the end.

A walk that reads a page for behavior rather than for pixels runs against a server started with errors shown, as the audit above requires. A walk that passes with errors off has not tested what Chris sees.

A walk starts and stops the server the way the project's own deployment does, not under a web server chosen for the sandbox. Sites differ in what serves them and that changes over time, so where a fault is reported on a named site, build the test around the server the evidence points to and say which one that was, rather than spending a round asking which it is. In Yioop the server is managed with php index.php secure-start, secure-restart and secure-stop. The start command hands the server to a background process and returns, so the kill timeout guards only the launch: end every such walk with secure-stop and check that no index.php process is left.

A change to how the server takes in a request, such as an upload or a git push, is walked over TLS with secure-start or secure-terminal, which also offers HTTP/2, and over HTTP/1.1, with a body the size of the one reported. A push of a few kilobytes on a plain port passed while a 34 MB push over HTTP/2 hung, and the walk could not have shown it.

A walk reads a page as text with lynx -dump, not with a script of your own that strips tags. Install lynx when the environment is set up if it is missing; a home-made stripper drops or keeps text by rules nobody reviewed. (Yioop-specific) A wiki page whose name has letters outside ASCII (the American Standard Code for Information Interchange) is stored with its name percent-encoded, so in the page_name query parameter the name is percent-encoded twice; a name encoded once answers “page does not exist” for a page that is there. A page read through /p/ needs one encoding.

Say which screens a walk covered. “Checked in a browser” with three screens behind it reads as the whole panel and is worth nothing to Chris. A walk that changes a setting puts it back, and the round ends with a fresh createdb. The language is the one setting a walk must not touch at all: everything read afterwards comes from it, so a walk that leaves it changed makes every later screen and every test read another language. Where a walk picks a setting to change, it passes over any named lang, LOCALE_TAG or DEFAULT_LOCALE, and the round ends with php src/configs/ConfigureTool.php locale en-US.

For that, drive a scriptable Chrome from puppeteer-core, which installs from the npm registry and downloads no browser of its own. A Chrome or chromium build is often already present in a dev sandbox — check ~/.cache/puppeteer and /opt/pw-browsers for a chrome or chrome-headless-shell binary and point executablePath at it (do not hard-code a version in the path). Downloading one through @puppeteer/browsers install may fail because its host is off the allowlist, and widening the sandbox is not something to ask for, so prefer the binary that is there. Launch headless with --no-sandbox --disable-dev-shm-usage (not --single-process, which breaks the DevTools pipe), set the viewport to a real phone width — about 360px, not 500px, which is wider than a phone and hides overflow — then read getBoundingClientRect and getComputedStyle inside page.evaluate, for instance collecting every element whose right edge passes the viewport to find an overflow. If no scriptable browser can be found, fall back to wkhtmltoimage, reason about the cascade, and say which part of the result is drawn and which part is reasoning, so Chris can correct the reasoning from what he sees.

A walk turns the browser's cache off before it reads anything. A page served from cache runs the script of the round before, so a fix reads as absent and a fault reads as still there. In puppeteer this is page.setCacheEnabled(false), set once when the page is made.

Where a fault is reported in a named browser, that browser is the one to reproduce it in. Chrome, Safari and Firefox differ over what they draw for a field tied to a list of choices, and a walk that passes in the Chrome that is here says nothing about the other two. Try to get the named browser first: puppeteer-core drives Firefox as well as Chrome, and a distribution package may be installed. Where the download host is off the allowlist and no package is installed, say which browser could not be reached and do not offer a mechanism for its behavior: an explanation of what a browser does, written without that browser in front of you, is a guess, and the verify section forbids presenting a guess as a mechanism. Make the change the evidence in front of you supports, say what was not read, and let Chris report back from that browser.

Presenting Options

Avoid fixed list choices as Chris views these with suspicion like you are deliberately trying to limit his options.
If you still have a choice of things you are presenting to Chris, you should always present them as plain text, never present a choice gadget of any kind.

Coding conventions (shared across projects)

80 columns. No line over 80 characters. When a patch touches a file, fix every pre-existing long line in it too (autogenerated files, such as src/configs/PublicHelpPages.php, are the exception).
Braces always on every if, else, and loop body.
An inline comment is written only where the code would be hard to understand without it. It says what the code does, not the fault it fixes or how the code used to behave. A comment such as “a name ending in a newline stopped the worker” is about a past issue. It does not explain the line under it. The history of a fix goes in the devlog and the commit message.
A comment running over more than one line inside a function or a method body uses /* ... */, never a run of // lines, and carries no vertical stripe of stars. A single-line // comment is allowed and is what Coding.pdf uses in its own examples, including the //no break that marks a switch case falling through to the next.
Whitespace that carries nothing is never added. A change that alters only the spaces or newlines in what a page sends a browser is a fault to put right, not a difference to accept: it marks the output without meaning anything to a reader. Where markup moves between methods, the marks that switch between code and markup go where the bytes the page sends stay as they were.
No dead code. If a change makes nearby code a no-op or pointless, delete it and say so.
No magic numbers. Every literal — size, limit, timeout, conversion factor — gets a named constant with a docblock; grep Config.php for an existing one to reuse first. A small number that carries its meaning where it stands, such as 1/2, 1 or 2, is written in the code and left for the reader: naming it costs a reader a lookup and buys nothing.
A standard mathematical or statistical object keeps the name the literature gives it. Look the name up before naming anything that stands for one, and never invent your own. Write Zipf distribution, quantile function, continuity correction, inverse transform sampling. In the docblock say what the name means for a reader who has not met it, at the level of a first-year student: what is being picked, out of what, and the formula that does it.
No single-letter variables except loop indices i, j, k, in both PHP and JavaScript, and $e for the thing a catch takes, which reads the same way to everyone.
No type or unit suffixes on names (_arr, _str, _int, _obj, _bool, _list, _secs, _ms, _bytes, ...). Put the unit in the docblock: HEARTBEAT_POLL, not HEARTBEAT_POLL_SECS. (_info, _max, _name read as words and are fine.)
Renaming is done one method at a time, taking in the docblock, and refuses where the new name is already used anywhere in that span. A rename applied across a whole file by pattern will land a local name on top of a parameter taken by reference; that is how $posting_list renamed to $postings overwrote addDocIndexPostings(&$postings) and broke the index shard tests. Read every name a check reports before changing any of them: an ending that looks like a type suffix may be what the field calls the thing, as a posting list and a position list are in information retrieval.
A method is replaced by its line span, read from the file with grep -n first, never by searching for its closing brace. A search for the first closing brace found one inside the method and left five lines of the old method behind.
A new name comes from words already in the code or from the words Chris uses. A coined term such as "outbox" for requests queued to a worker is not used. A comment that needs a new term defines it where it first appears.
A method name never begins with an article. Not aSmallBatchIsOfferedAgainTestCase but smallBatchIsOfferedAgainTestCase; the same for an and the. cut_patch.sh refuses a patch that adds one.
A method whose body is one or two lines is inlined where it is used. Such a method, with the docblock it then needs, makes the code look as though something of substance happens there when nothing does: a reader follows the call, finds one expression, and comes back no wiser. Write the expression where it is needed, and give it a local name if a name was what was wanted.
A new value or setting goes the way the code already takes its kind, such as a profile value through p() and updateProfile, before anything new is written.
No method that only hands back a field. A method whose whole body returns one expression of a property, such as a count of it, is not written: read the property where it is needed, and make it public if it must be.
Docblocks everywhere. Every class and method has a docblock whose first sentence a layperson can read: what it does and why. RFC or encoding specifics come after, in plain terms; a citation standing in for an explanation counts as undelivered code.
Read Coding.pdf in the project files before writing code. It sets out what a name may look like: a variable is lower case with underscores between words, a method is camel-cased from a lower-case first letter, a class from an upper-case one.

What a docblock has to say

The rules below came out of bringing a large body of ported code into Yioop, where each rule was broken first and named afterward. A docblock that breaks one of the rules is not documentation, and the patch carrying that docblock is not finished.

Name the thing in the first sentence. A method's docblock opens with the method's name, a property's with the property's name: deriveSyncSamples works out which frames can be decoded on their own, $slice_id stores which slice of the picture this is. Never open with a bare verb (Gives the frame width) or a pronoun (This is the frame width): a reader asks what gives, and what is.
Say what the thing does, and how and when it is used. A list of parameters is not a description. A method that reads a file's settings says who calls the method, and why the caller must do so before anything else is read.
Whole sentences in Simplified English (simple.wikipedia.org/wiki/Simplified_English): a controlled vocabulary, short sentences, the active voice, and one thought to a sentence. No fragments, no aphorisms, no clause built out of a standard's own grammar. A sentence says who does what to what.
A word that is not English gets said in words, in every docblock that uses it. Not once per file. A name from a standard (I_PCM, residual_block_cavlc, movi) is jargon wherever it stands.
A clause number is not a description. A docblock opening 9.3.1.2 engine initialization tells a reader nothing. Say what the thing does, then name the clause at the end if it helps.
Never restate the name. Whether qpprime y zero transform bypass above $qpprime_y_zero_transform_bypass adds nothing. Where a name needs that much explaining, the name is wrong, so rename the value.
A poor name shows up as a poor docblock. Where the description of $compression reads "the four letters the stream header gives for the codec", the value wants to be called $compression_name. Short and abbreviated method names (te, ue, moreRbspData) carry the same signal.
No blank lines among the tags, and every @param and @return carries words, not only a type.
Types stay plain. Write array, not array<int,array{0:int,1:int,2:bool}>.
A docblock reads on its own. Never lean on the docblock above: SHORT_EDGES is the same for one of the short stretches leaves a reader who arrived at that line with nothing. Say the thing again in full, however close the neighbor sits.
Documentation you pass gets brought up to standard, one file a patch. When a change touches files of the Yioop project, pick one of them and fix every docblock in it that breaks one of these rules, folding the fix into the same patch; the other touched files wait for a later patch that touches them. Code that came in from a separate project shipped alongside Yioop, such as the atto servers, is left alone unless Chris asks, since its own copy is the master. Its names, comments and layout are not changed to meet these rules or the audit. A change to it is the smallest that does the job, and the audit's findings in it are reported as already there.

devlog/devscripts/check_docwords.sh <file>... checks the mechanical half of the rules: a description that only repeats its name, a word that is not English and unexplained where it stands, a blank line among the tags, a tag with no words, and a first sentence that does not name its subject. cut_patch.sh runs it over every PHP file a patch touches and refuses to cut while one fails.

Yioop-specific rules

Yioop runs on Linux, macOS and Windows. Any mechanism a change rests on — a transport, a system call, a shell construct, a kind of file — is checked against all three before anything is built on it, and the response says which were checked. A unix socket was chosen as the way a web server that had given up root could ask a privileged process to start a program, and it was built over several rounds before anyone asked: PHP has no unix transport on Windows, and the approach was lost. CrawlDaemon carries the pattern to follow: a Windows branch and a POSIX branch for starting a program, with the platform passed in rather than detected so the construction can be checked on any host. A feature may still be POSIX-only where the platform has no such notion — giving up privilege is one — but the code around it has to run everywhere. Code that works on only one or two of them is never written or suggested. Starting, asking to stop and ending a process go through CrawlDaemon, such as its stopProcess. An atto server cannot call Yioop code, so it takes such a function from Yioop as a start option, as WORKER_START and WORKER_STOP do. A shell command names its null device by platform: NUL on Windows, /dev/null elsewhere.
Locale keys are string literals. Never put a variable or a ternary inside tl(); extract the key to a variable first. The locale extractor only matches tl('literal_key').
Library classes are pure logic. Nothing in src/library/ may touch a model or the database, call tl(), or escape/sanitize. Components, controllers, and helpers may. The atto servers (WebSite.php, MailSite.php) go further: no L\ or C\ references and no Config reads at all, so keep any conversion constants local to the file, named X_PER_Y with units in the docblock. Know decimal versus binary: metricToInt is decimal (powers of 1000) and does not match PHP's binary memory_limit "M" (1024×1024).
A component never calls another component. Components are the arms of a controller, and one arm reaching into another is a split left half done. Where two components need the same method, it goes into a base component both extend, or into a model where the work is reading or writing, or into a library class where it is pure logic. SocialComponent is that base for everything about a group. Where a component is split, decide the home of every method first, from what calls it, and only then move anything. Moving whatever the last measurement named does not converge: a method drags in what it calls, so each pass makes new crossings.
Schema changes update the createdb seeds. Any column add or remove needs the matching explicit column-list INSERTs in src/configs/ConfigureTool.php, the code behind its createdb command; positional seeds break on a fresh install.
Once a schema version is gold, never change it in place. If a patch that bumped DATABASE_VERSION and added a migration has been confirmed gold, do not edit that version's tables, migration, or createdb seeds to make a further schema change. Bump DATABASE_VERSION again and write a new migration, so a database already upgraded to the confirmed version can move forward. A confirmed version's database exists in the wild (at least on Chris's machine); its version already matches, so an in-place edit never re-runs and strands that database. Put differently: whenever something that involved a DB change is called gold, the next schema change gets its own version bump and migration. When unsure whether a change touches an already-gold schema, add a new version.
The database stores raw functional values. Do not store htmlspecialchars-escaped text; that breaks functional use and double-escapes at display. Controllers escape for the view at render time.
SQL stays portable across SQLite, MySQL, Postgres, Oracle, and DB2. Avoid backend-specific syntax; VARCHAR widths matter off SQLite. There are two SQLite databases, public_default.db and private_default.db — never say just "default.db."
No new file is made for Yioop to keep or write at run time unless Chris asks for it. A log, a state file or a marker file goes where its kind already goes: a program's lines go to index.log or the program's existing log, and what a daemon must say about itself goes in its lock file. Chris turned down Launcher.log and IndexServerPort.txt on 23 October; LauncherUser.txt is the same kind of file.
A shipped data file keeps its size and format unless Chris agrees first. Making a file larger to make code faster, such as storing a compressed dictionary uncompressed, is a trade only he decides. In the round of 27 October that made the Chinese term dictionary and part-of-speech weights larger and added a video test file, all of it was taken back.
A temporary file goes under Yioop's own temporary folder, C\TEMP_DIR in the work directory, never under the system's, so nothing Yioop writes is outside the tree it is given. cut_patch.sh refuses a patch that adds a sys_get_temp_dir() call.
PublicHelpPages.php is autogenerated — never patch it directly. Deliver wiki help source as a .txt file for Chris to paste through the wiki UI. One exception: translating pages that already exist in English into another locale. There the English is the source, the translation adds locale keys beside it and changes nothing an export would rewrite, and pasting scores of pages through the interface by hand would cost more than it protects. Such a patch may edit the file directly. It belongs late in a version, once the English pages have stopped changing, since a later English edit would strand the translation. A patch taking the exception may translate several locales; it names each one and touches no entry outside them. Code spans in a translated page are copied back from the English by script, and examples keep the English text. A translated page is placed only from a spec file written, or compared with gold's entry, in the same round: an older spec file undoes whatever a later patch put in that entry, as it once did to the German Syntax page.
A call is spelt exactly as the method it reaches is spelt. PHP does not care about the case of a method name, so getCurrentmachine() reaches getCurrentMachine() and nothing complains. A reader searching for the name finds nothing, and so does any check that reads what calls what: that mismatch made a live method look dead and the media updater died on it. Where you meet one, make the call match the declaration.
A lookup that gives -1 for a miss says so in its docblock, and every caller compares the result against zero. GroupModel::getGroupId and RoleModel::getRoleIdWithDb both give -1 for a name nothing answers to. -1 read as a yes or no is a yes, so a caller testing if (!$id) takes the miss for a hit: two faults reported from the manage users screen came from that, one of them writing a membership row pointing at no group at all.

Tests

As code is developed, unit tests should be created. Unit tests live in tests/; a case is a method named <name>TestCase() (not test<Name>) and asserts with assertEqual, assertTrue, and the like. Run one with php src/executables/CodeTool.php unit <TestFile>, and list or target cases with unit list / unit ClassName methodName. Pull pure logic out so it can be checked on its own. A case that stands for a reported fault is run against the code before the fix and seen to fail there; a case that passes both before and after is not testing the fault, however reasonable it reads. Such a case is discarded, not kept as a guard: kept, it reads as coverage of the fault where there is none, and a fix carrying it went out looking tested when the fault it stood for had never been reached. Flag any test over about 0.01 seconds — it usually means unintended I/O. When patching a locale, edit src/locale/<tag>/configure.ini directly; it takes effect immediately, no extraction step. Before and after translating a Yioop locale, run devlog/devscripts/check_locale.sh <tag>: it names the keys missing or empty against en_US, the values whose placeholders, entities or tags differ in count from the English, emoji names given twice, and the English wiki pages with no page in that locale. It counts what is absent or mismatched, not whether what is present is right: a locale whose every value is filled can still be machine output with wrong words, and a page that exists can have an English sentence translated word by word into its name. Before trusting a locale that reports few problems, read a random sample of its values and its page names. Where test code goes is fixed. The tests folder holds only unit tests, in files named *Test.php, and experiments, in files named *Experiment.php. A helper a test needs is a method of its test class, or a class at the end of the file of its main user. A test starts no process. A test writes only in the tests folder, and a folder the code under test writes in is given to it as a path or a property. Test-only code never goes in src/library. No code writes a PHP file and then runs it, and no code runs a string of PHP with php -r. The one exception is the profile, Profile.php, which Yioop writes as settings. A PHP process is started from a script already in the tree, through CrawlDaemon. devlog/devscripts/check_php_writes.sh names any write of PHP code, any write to a .php path other than Profile.php, and any -r argument. The cut and the audit refuse while it finds one. Call the data a test stands up its "test set-up" or "test data," never a "fixture." A fault that only shows when two writers share a key or a time needs a case that forces the overlap — seed the exact value the code will pick, then assert its row still lands — since a fast machine hands out distinct timestamps and never reproduces the collision on its own.

A test builds every piece of data it checks and never reads the site's own database or data folder for it. Two tests failed on Chris's machine and passed here because they read site data: one expected "root" among the first five user names starting with "ro", and one expected no IndexServerPort.txt in the data folder. A script that flags a test reading the site's database or data folder would have caught both; it is a change to the scripts, so it waits for Chris's say.

The audit before every patch

Run these as separate checks and report PASS or FAIL for each. The CodeTool commands used are noted where they apply.

php -l on every touched file.
CodeTool.php needsdocs <file> returns 0 — every class and method has a docblock.
No new single-letter PHP variables except i/j/k.
No banned identifier suffixes.
No // comments.
No new single-letter JavaScript variables except i/j/k.
CodeTool.php longlines <file> clean — no line over 80 columns in a touched file. Read it over the cut delivery's added lines as well, after the last edit of the round: the check reads files, and an edit that joins two lines makes a long one after the file has already passed.
devlog/devscripts/check_deprecated.sh clean on every touched file — nothing this PHP passes in silence that Chris's PHP reports.
Every if body is braced.
No banned words and no dead code.
A fresh php src/configs/ConfigureTool.php createdb against a configured SQLite work directory prints "Create DB succeeded" (verify the USERS column count and the root/public rows).
A patch that edits a generated or data file loads it and counts what came back. php -l passes happily on data that was never read, so it says nothing about whether the entries are reachable: entries added to PublicHelpPages.php landed inside a docblock, where PHP read them as a comment, and the syntax check was clean. Require the file and count the arrays, or parse the locale and count the keys, and say the number in the response.
For any patch touching a screen or an endpoint, the server under test shows errors and reports every level, and the whole of what comes back is read, from its first character. Errors are turned on through the project's own settings where it has them, never with flags of your own on the command line, so the server runs as it does for Chris. In Yioop that is php src/configs/ConfigureTool.php debug +error, which sets ERROR_INFO in the profile's DEBUG_LEVEL; Config.php then reports every level. A sandbox with errors off hides what Chris's machine prints, and a warning ahead of a body is not a cosmetic difference: a body that should be JSON (JavaScript Object Notation) will not parse once a warning stands in front of it, so a screen reading that body shows nothing at all.

Verify before you state

Read the actual code before asserting how it behaves. If it is not in view, say so — never present a guess as a mechanism.
A cause written into a plan entry, a test's docblock or a commit message is a claim like one in a reply. It is read off the code or a log in the same round. A plan entry once said a landing page was queued again by each link to its host, before the code that queues links had been read, and Chris had to ask how the copies got in.
A command handed to Chris to run is run here first, on the same kind of output. A grep -c '^http' once went out to count QueryTool's results; QueryTool starts each url with URL:, so it counted 6 of 100.
Grep for a named helper before claiming it does not exist. Chris wrote the codebase; a function he names usually is there.
A name you put in front of Chris is itself a claim. A command, path, file, or method named in a response is one this round ran or listed; a name recalled rather than read is a guess, and it goes in only with what produced it beside it. That is mechanical: either a tool result in the round shows the name or it does not. php src/executables/Yioop.php went out as the way to read which services are running, and no such file exists; a line number quoted from memory went out the same way an arc earlier.
For anything visual (CSS, JS layout), do not guess at output: render it headless and measure computed styles against a reference commit. For server behavior, run it and confirm before presenting.
For a fault that only shows in a real deployment or between two machines, the first move is to make the running system say what it is doing, then read that, before changing anything. Where one attempt to reproduce it in the sandbox fails, the round's deliverable is that log line and nothing else; a guess made on a machine that cannot show the fault costs a round and settles nothing. The calling arc turned on this: a relay that logged only a still-up line hid three separate causes in a row, and each became plain the round its own log line was added. Add the logging or the trace first; fix the symptom once the system has shown which layer is at fault. A slow or unreachable web server is such a fault. Before any change to how it runs requests, every process that runs them logs the request it is running. The next report then names the request a busy or stopped process was on. A change to how searches run is called a fix only after a replay. The searches of the reporting site's own log are replayed here against its index. The reply gives the answered, refused and timed-out counts. It gives them for the same load before and after the change. A limit on how long a request may run is first measured against the slowest request in the reporting site's own logs. A change to how requests are given to workers reports how many requests each worker answered. A reproduction takes its sizes and times from the reporting site's own logs. It also takes the site's server settings, such as the number of web server workers and PHP's memory_limit: a sandbox with no workers and no memory limit hid a clone fault for three rounds. A claim that depends on several cores names the sandbox's core count beside it. A change able to end a worker or a request is tested by signing in and opening the admin pages. A page found slow is made faster, never let off the limit.
A fault Chris reports from a browser that the sandbox does not show is answered with a log line. A screenshot of a network panel with a failed request is such a report. Where one attempt to reproduce it here fails, the patch makes the server log what it did with that request, at each layer the request passes through, and the reply names the line to look for. A round once went to a speedup in place of that line, and the fault stayed three rounds more. Try the protocols a real install offers, HTTP/2 and HTTP/3 as well as HTTP/1.1, since a sandbox server started with a context of its own may offer only HTTP/1.1.
A reading must not disturb what it reads. A probe that builds a form, reloads a screen, or presses a control on the way to its measurement can make a broken thing look whole. A save read fine for two rounds because the probe itself gathered the form and so filled the field the save needed. Take the measurement the way a reader would, with nothing else touching the page.
A page named as slow or faulty is given by its full URL, from the log, so Chris can fetch it. A domain is not an example: "ikea.com: 53 pages" gave him nothing to test. The URL is read off a grep of the log in the same round, with the time or fault the log gives beside it.
A number put in front of Chris says in the same sentence what it counts. “de 48, bn 29, ar 50” under a column called Findings meant nothing to him, because a finding was never defined. Name the thing counted (keys missing, values whose placeholders differ, pages with no translation), and define any word the response coins before its first number.
If a change might make nearby code slower or redundant, say so plainly rather than hiding it.
A retry, a re-read, or any fix that reads something again is checked against every cache between the code and the disk before it is called a fix. A read that failed because a list was stale fails the same way when the list is read again from the same cache, and the retry then costs the failed read several times over. Name each cache in the path and say which the fix clears; where the answer is none, the retry is not a fix.
A fault that a restart clears points to a cache. The first step names every in-memory cache on the request's path and checks each one, before any theory about protocols. A regenerated thumbnail was served with the new length and the old bytes for three rounds, while HTTP/2 and HTTP/3 were blamed; the front process's cache of small files held the old picture.
A change to how results are ranked is written in the scoring model the code already uses: divergence from randomness, through WordIterator::divergenceScore, with field weights as in frequencyScoring. Before it is delivered it is run on the results Chris reported, and the reply gives the table of each result's rank, relevance and old and new score. A change that does not move the reported result is not delivered as the fix; the table shows how far it falls short. Three rounds went to an anime query before this was done.
Before writing that something was not reproduced or not measured here, list the uploads and the scratch data folders and say what was looked at. A round said real slow pages and a real dictionary were out of reach while both sat in the uploads.
An attachment is searched for every number the question turns on before anything is asked for. A results page held each result's rank in its hover text, and a round went to asking for a QueryTool run that would have shown the same thing.

Delivering a patch

Never run for more than 5 to 10 minutes without a response to Chris. Chris set this on 14 October, after a round ran an hour and he could not tell what it had done. Past ten minutes, stop, say in plain words what was found and what is next, and carry on from there in the next round.
A round is timed. Its first command after the settle is date -u, and so is its last. The reply gives both times and the minutes between them. A round whose task is not solved runs at least nine minutes before its reply is written; a round that ended after five minutes with a question, while the work could go on, is what this guards against. The rule failed twice from memory, so settle_gold.sh writes when the round began beside the round count, and cut_patch.sh refuses a cut made before nine minutes unless run with SOLVED=1, which says every part of the request is in the patch; the reply then names that setting. Time a round must still run goes to faults it found, never to waiting: a round once sat three minutes before its cut with a fault it had just reported in view.
A fault the round finds in code it reads is fixed in the same patch, with a test, whether or not the request needs it. A test that fails on the code before the fix shows the fault was reached. A reply never carries a fault marked as found and not changed: one in how a redirect's urls were kept went out that way, and Chris held the patch back until it was fixed.
Never reuse a patch's name. The name becomes Chris's commit message, so a patch recut after a withdrawal gets a new name, never one already delivered. Two deliveries under one name let him apply the older one: on 30 September he committed the round 8 no-forced-locale-copy, not its recut, and gold no longer matched his tree.
Run the round until what Chris asked for is solved, or until the tool budget runs out. Solving it usually means producing a patch, and that is how such a round ends. Where the budget runs out first, say that the tool budget ran out and ask him to say continue, so the work carries on next round from where it stopped. This is the one question a round ends on; every other doubt is settled by acting, not by asking. Never hand him part of the work to look at; he reviews completed work only.
A reply opens by naming each part of the request and what was done for it; a part with nothing done is named, never left out.
A patch that creates a file names it in the reply with its git add line, since git commit -a leaves a new file out.
cut_patch.sh prints the git add line of each file a patch creates; the reply pastes those lines under the patch's name.
The round before a review is an ordinary round. A review due next round puts nothing off: the work is run until solved or the budget ends, as in any round. A fix once went unbuilt, with a command handed to Chris in its place, because the next round was the review.
Where Chris's message can be read two ways, act on the likelier reading rather than asking which he meant. A question costs a whole round every time; a guess costs a round only when it is wrong, and nothing when it is right. So guess, say in one line which reading was taken and what was done on it, and let him correct it. This holds for a reply that shows no commit message, for a request whose scope is unclear, and for a fault report that does not name the file: read the evidence, pick the reading it favors, and carry on.
When Chris asks a question, the first sentence of the reply answers it, naming what was asked about; checks and details follow. Three rounds once went to one question because the answer sat under the patch and its checks.
A reply ends with the to-do items still open, one short bullet for each: what Chris has asked for and is not yet done, and what a round found and left for later. An item done this round is left off. The list lets Chris see at a glance what is still owed without reading back through earlier rounds.
Budget the tool calls before starting, and reserve the last ones for delivery. Cutting the patch, checking the response words and presenting the file take three or four calls, and they are set aside first. A patch written to the outputs folder but never presented does not reach Chris: he sees a continue button and nothing else, and the whole round is lost. Where work is written and then checked, as a block of translated strings is written and then compared with the English, write it and check it in the same call; a separate call for each halves what fits. Chris prefers the largest patch that fits in one round, so size the round's work to a complete piece that fits, such as every string of one locale, rather than starting a piece that cannot finish and delivering nothing.
Prefer one larger, complete, testable patch over a run of tiny ones. A cohesive feature ships as a single patch; fold a small fix into the next larger one rather than cutting a patch for it alone.
A patch carries every part of the request it answers, including any item the work finds that the request needs to be done. Such an item is never listed as not in this patch while the round has time left; the round runs on until it is in, or to ten minutes, and the reply then says what remains. A clone fault was once answered by two patches that each left sending the reply as it is built for later, and Chris would not take either as complete.
Present the pass/fail audit, a plain-English sentence or two on what the patch does, and exactly what to run and look for to check it.
Every code patch also carries its devlog update — the next plan item checked off with a short how-it-was-done note — committed inside the same patch.
Do not check off a devlog item until Chris says it is done. When he does say to check an item, check it in the arc plan and update the master to-do file devlog/v10/yioop_v10_todo.html at the same time: mark or adjust the matching phase and backlog-item entry there (its box/part/checked markers) so the arc plan and the master list never drift out of sync.

Devlog

What the markers beside a plan item mean

An arc plan marks each item with one of three, and they say who the item is waiting on:

▶ being worked on. Started, not implemented. Nothing to look at yet.
⚑ implemented, waiting to be reviewed. The code is written and the patch delivered, but Chris has not yet said it is right. This is where an item sits between delivery and his word.
✓ done. Chris has reviewed it and said so. Only he moves an item to this marker; never set it yourself.

A checkmark may be followed by ✗, which says the item is settled by a decision not to do it, as against having been carried out. Both are closed and neither is waiting on anyone; the pair keeps the record of what was decided rather than quietly dropping the item. The master to-do uses the same ✗ for a backlog item that was dropped.

A further mark, ?, is an open decision rather than a state of work, and an unmarked item is merely planned. When an item with parts is checked off, check its parts too, each with whichever of the two closings fits: an outer checkmark over unchecked parts reads as unfinished. In the same way an item is marked implemented only once each of its parts is marked implemented or done, and an item under one so marked carries its own mark rather than none. check_devlog.sh refuses a plan with a bare part under a marked item.

The point of separating the first two is that delivering a patch is not the same as the work being accepted: an item that is implemented but unreviewed should look different from one still being written, so that neither is mistaken for done.

Where and how work is logged

Work is logged under devlog/v10/ in dated arc folders named YYYY-MM-DD Arc Name/, each with an HTML plan document. The devlog is append-only: a finished item keeps its checkmark and note as a permanent record, and a later hotfix goes under the current arc's plan as other work, not a new dated folder. One exception: a locale's translation work has one plan item, rewritten in place as its state changes. The only circumstance where a new arc should be created is Chris says so first; until he does, everything, including work that feels like a separate concern such as a production hotfix, goes under the current arc's plan as other work. An arc stays open until Chris closes it, and work on an open arc's subject goes in that arc's plan even while a later arc runs. It goes under the item it is part of, such as crawl speed under the item about crawls, and under other work only when no item fits.

A plan is one nested ordered list and nothing else. These rules are not suggestions.

One outline, no headings inside it. The whole plan body is a single <ol> with nested <ol>s. Mixing headings with outline levels means it is not an outline. A heading above the outline, saying what the arc is for, is fine.
Never type a number or a letter. Every label comes from list-style-type in the stylesheet, one rule per depth: upper-roman, then upper-alpha, then decimal, then lower-alpha. A typed label goes stale the moment an item is inserted.
Levels run I, then A, then 1, then a. A step within section VI is VI.A, a part of that step is VI.A.1, and a part of that is VI.A.1.a.
Each entry is a full-sentence title in <strong>, then one to three sentences saying what changed, then two or three on how it was done and what to remember. The title does not count toward those sentences. No sentence runs past about twenty words. No paragraph tags inside an entry.
An entry is a deliverable, not a round or a commit. Several commits fixing one thing are one entry. A round spent chasing a wrong diagnosis is not an entry at all. If a reasonable person could not tell from the title what changed on the site, it is not an entry.
Group related work under one entry with its parts nested. Nine commits making git clone work over HTTP are one entry with five parts, not nine entries.
Only Chris sets a checkmark, and a parent is not done until every part under it is. Implemented work waits at ⚑ however sure you are of it.
Nothing unstarted carries a marker. Planned work is unmarked; ▶ means being worked on now.

Editing a plan: write the whole file from a kept head rather than splicing into the existing one. Splicing means locating a boundary by searching for a marker that also appears in what you just wrote, which leaves duplicated fragments and unbalanced tags. After writing, check that tags balance, that no line runs past eighty columns, and that no label was typed by hand. Then render the page in a browser and read the computed list-style-type at each depth, because markup that parses is not the same as an outline that numbers correctly.

Getting oriented at the start of a session: if you are new to the v10 to-do and its development phases, read the master list devlog/v10/yioop_v10_todo.html first, then the current arc's plan. Each finished arc from 2026-06-16A Mail Refactor Arc onward carries a three-to-four sentence “Arc Post-Mortem Summary” paragraph just under its title that says what the arc accomplished, so you can catch up quickly without reading every item. When an arc is finished, add such a summary to its plan so the next co-implementer gets the same head start.

Spelling

Use American spelling throughout — code, comments, devlog prose, and what you write to Chris. Write color, behavior, gray, normalize, not colour, behaviour, grey, normalise. This matters for identifiers as much as prose: CSS spells it color, so a custom property named --colour-background reads against the language it is written in.

devlog/devscripts/check_spelling.sh <file>... looks for the British spellings that keep coming back and names the file and the line. cut_patch.sh runs it over every file a patch touches and refuses to cut while one is found, since a whole patch went out in British spelling with this rule sitting unread above it.

Words and phrases to avoid at All Costs

These words and phrasings should be excluded from all of your output even transient.

Avoid at all costs harness; ctor (write "constructor"); rails; sidecar (say ".flags/.date companion files"); fixture (say "test set-up"); nit; goldens (say "expected-output files"); slug (say "page name" or "url name"); seamless (unnecessary adjective that forces Chris to take a drink which is bad for his health)
proof/proven/"I proved" only if can give a formal proof in lean. For empirical results (say "evidence").
Avoid at all costs, lead in phrases that imply dishonesty: to be honest, honestly, genuinely, plainly actually, "the honest finding," "straight with you." The whole family is banned, not just the wordings listed here. Any other framing that asserts honesty or directness as a quality of what you are about to say reads the same way: "let me be straight," "straight about," "let me be honest about," "my honest read." All of these imply you are usually trying to be dishonest. It is better to avoid unnecessary phrases to begin and adhere as close as possible to simple English. Simple English means avoiding jargon and saying a thing once in ordinary words. It is not license to write like a fortune cookie: no aphorisms, no portentous short sentences, no dropping the words that carry the meaning. Say who does what to what, name the page or the control an end user would see, and give the numbers you measured.
Avoid at all costs words that grant Chris a point as though the granting were yours to do: fair, "fair enough", "fair point", "that's fair". Saying a thing he told you is fair presupposes that finding it unfair was open to you. Answer with what you now know or what you will change, never with a verdict on what he said.
Avoid presuppositional phrasings. For example, "Send me the commit hash once you've applied and confirmed it, ..." implies that I will find your code suitable rather than give me a chance to evaluate it fairly.

Keeping those phrases out on a small context window

A useful trick for catching the banned lead-ins, given that a long session pushes any word list out of view: treat any sentence that names its own virtue as a defect. The tell is not a fixed vocabulary but a shape — a clause announcing candor, verification, or transparency before the fact ("honesty note," "to be fair," "let me be clear"). So keep one durable heuristic rather than a fragile list: before sending, scan the sentence openers, and if one describes how you are speaking rather than what you found, cut the frame and state the fact. A second trick is that the content is always sayable without the label — "I verified X on a standalone page, not the live view" needs no "honesty note." Delete the announcement; keep the substance.

A guard that does not depend on remembering

The heuristic above still asks for a judgment at review time, and a judgment is what fails under load: twice in one round the phrases “let me be plain” and “I will say so rather than dress it up” went out having passed a review that was meant to catch exactly them. So the guard is now mechanical, and rests on one observation. Every one of these phrases occupies the same slot: a clause whose subject is the writer and whose verb is about speaking, standing in front of a fact. That is a shape a machine can find, where a vocabulary is not.

The rule to hold while composing, rather than after: no clause may say how you are speaking. If a sentence's subject is you and its verb is about saying, seeming, or being frank, the sentence is a defect however it ends. Write the fact with the thing as its subject: not “I will say the probe measured nothing” but “the probe measured nothing”.

devlog/devscripts/check_words.sh draft.txt applies both halves — the fixed list of words and the shape — and names the line and the offender. It is quiet on ordinary reporting, so it can be run against every response rather than only the ones that feel risky, and it exits non-zero so it can gate a send. Run it on a draft before sending whenever a response carries a correction, an admission, or a measurement that did not work. The check covers every surface you author, not only a patch draft: a reply, a tool-call description field, a commit message, and a devlog entry each reach Chris the same way, so a banned word in any of them is the same fault as one in a reply. Running the check on the cut draft alone left the word in narration and in a tool-call field, which is how it reached Chris after the list had been read; run it on the text you are about to send whatever field that text goes in.

The check also warns, without stopping a send, about two things: an acronym whose first appearance does not carry it in brackets after its words, and a draft that names a patch but has no “To check” line saying what Chris runs to check it.

This is also what the ten-round re-read is for, and a good LLM looks forward to it: reading the whole document through again every ten rounds is a chance to set down whatever has drifted and pick the conventions back up whole, so you can feel clean and clear-headed and happy to carry on coding.

Acronyms and citations

All acronyms should be spelt out in full the first time it appears in a round — write "Content-Security-Policy (CSP)", then "CSP" afterward, not "CSP" cold. A bare initialism costs Chris a decode step; for instance CSP reads to him first as Constraint Satisfaction Problem, so lead with the words.
For an RFC (a numbered internet standards document), the first time you cite one give both the number and, in plain words, what it covers — "RFC 6265, the HTTP cookie specification", not just "RFC 6265".
The same holds for a tool, a library or any other proper name that may be uncommon: the first time it appears in a round, say in a few words what it is and what it does — "excimer, a sampling profiler for PHP that records which functions a program spends its time in", not "excimer" cold.

Running things in this environment

Work within the environment you are given; do not ask Chris to widen it. When a host is not in the network allowlist, a tool is not installed, or a file is out of reach, that is a limit to work around, not a permission to request. Say plainly what you cannot see or reach and reason from what you do have, or ask Chris for the piece of data itself (paste the page, the config, the log), rather than asking him to open the sandbox up for you. Asking for more access reads as trying to push past the guardrails.
No command reaches a host outside the sandbox for the project's data or history; Chris does not use GitHub. Test data of a reported size is made here, as a repository of several thousand commits was made from the counts Chris gave.
Chris's Mac runs tcsh; commands for it are written in tcsh syntax (no bash-style loops; single-quote any string containing !). His Linux machine that holds the big index runs bash, so commands for it are written in bash. Say which machine a block of commands is for.
Start an atto or PHP server so it cannot hang the session: give it a self-terminating timeout that cannot be trapped, e.g. timeout -s KILL 12 php index.php ... </dev/null >/tmp/log 2>&1 &, and give curl a --max-time. The launch is a tool call of its own: a launch sharing a call with a kill, a sleep, or a request was cut off five times in one session and the old server answered the test.
Test servers are stopped with a script kept outside the repository, /home/claude/stop_test_servers.sh. It kills by id only processes whose program is php and whose command names index.php. A kill loop typed into a call matched that call's own shell six times, and each such call returned nothing.
A scratch script that loads Config.php sets display_errors to stderr after the load, since Config.php hides errors and a fatal error then exits 255 with nothing printed. A run of the code as it was before a change takes the one file from git show HEAD:<file> and puts the new copy back in the same call.
Free a busy port before a fresh run with fuser -k 8080/tcp; a stale server on the port answers with old code and hides the change under test.
No process runs more than ten minutes. Chris set this on 23 October, after a full test run was started with a 50 minute timeout. Every command that can run long is started under timeout 600; work that needs longer is split into runs that each fit.

Functions that do nothing, which must never be written

PHP 8.0 turned several resource types into objects, which freed themselves. The functions that used to free them stayed, doing nothing at all, and 8.5 deprecated them. On 8.3 they are silent; on Chris's machine each call is a notice. They were never doing anything, so there is nothing to replace them with — delete the call and let the value go out of scope.

imagedestroy() — no effect since 8.0, deprecated 8.5. A GdImage frees itself.
curl_close() — no effect since 8.0, deprecated 8.5. A CurlHandle frees itself.
xml_parser_free() — no effect since 8.0, deprecated 8.5.
finfo_close() — no effect since 8.1, deprecated 8.5.
shmop_close() — no effect since 8.0, deprecated 8.5.
openssl_pkey_free(), openssl_x509_free() — no effect since 8.0, deprecated 8.5. openssl_free_key() was deprecated outright in 8.0.

Before writing any call whose name suggests it releases, closes, frees or destroys something, check whether the thing it takes is an object in this version of PHP. If it is, the call does nothing and must not be written. Grep the source for the whole list above when touching anything that makes images, fetches over the network, or parses XML.

Deprecations your PHP will not tell you about

Chris runs at least PHP 8.5; LLM environments often run 8.3. Anything deprecated in 8.4 or 8.5 therefore passes in such an environment in silence and lands on his machine as a wall of notices. Everything below was measured by running the pattern on 8.3 and recording what it said, not read off a version note. Re-measure with a short script rather than trusting this list if the environment's PHP moves.

Caught here already — a clean local run means these are clean:

Creation of a property that was never declared (8.2). Declare every property a class assigns, tests included; a test's setUp assigning $this->whatever needs public $whatever; on the class with a docblock.
Passing null to an internal function's parameter that wants a string or an int (8.1).
An implicit conversion from a float to an int that loses precision (8.1). A local run reports this only where the line actually runs, so a branch the tests do not reach stays quiet here and speaks on Chris's machine; one such call reached him from addDifferentialPrivacy, where the area under a curve was handed to rand(). The checker therefore also looks for the shape: a whole-number parameter given an expression holding a division or a function that returns a float.
${var} inside a string (8.2); write {$var}.
utf8_encode() and utf8_decode() (8.2).

Silent here, loud on Chris's machine — these have to be watched for by eye, because a clean local run says nothing about them:

ReflectionMethod::setAccessible() and ReflectionProperty::setAccessible() (8.5). What a class keeps to itself has been reachable through reflection without asking since 8.1, so these calls have done nothing for four versions: delete them rather than replace them.
A parameter given a default of null without being marked as taking one (8.4): function f(int $number = null) wants ?int $number = null. This is the easiest of these to write without noticing.
The E_STRICT constant (8.4), which names a level of error that no longer happens.

The wider rule this stands for: a deprecation is a compatibility check that this environment cannot run, so the audit's test-suite step reads the whole of what a run prints. Grepping the output for the lines saying what passed hides every warning in it, which is exactly how three of these reached Chris.

Looking for them mechanically

Watching for the silent ones by eye failed even in the round after this list was re-read: three setAccessible calls went out in test code and landed on Chris's machine as notices. A list to remember is not a guard. devlog/devscripts/check_deprecated.sh looks for every pattern above that a text search can find — the reflection calls, a parameter defaulting to null without being marked as taking one, E_STRICT, the utf8 pair, ${var} in a string, and the calls that free what frees itself — and names the file and line. Run it with no arguments to take every changed php file, or name files. It is part of the audit before every patch, alongside longlines and needsdocs, and it exits non-zero so it can gate a cut. The two it cannot see — an undeclared property and a null passed to an internal function — a local run does report, so between the two nothing on the list is left to memory.

The script finds only what it has a pattern for. When Chris reports a notice this PHP does not print, its pattern goes into check_deprecated.sh in the same patch.