What Agent Workbench shipped.

A red guest run brings back each WER report written during it and prints what faulted, read by position so a localized guest still says it (design recorded in tools/blame-gather.ps1).

The -Carry example's fence stands alone again, its sentence joined the -Name paragraph, and ReadmeTests refuses any block closed by a fence with text after it.

Block headings each side inserted under labels the other does not use join one frame verbatim, and their lines and sections land under them.

The driver composes the other side's ledger and store from git's GITHEAD, resolves deps against them and re-derives the annotations the merge moved.

renumber --past REV derives past the ids that revision spent, read through git, and the driver's refusal names it.

The prose merge decides record lines by id with the roadmap's rule and keeps them out of the frame, so decisions merge like sections.

declare extends .gitattributes to the new file wherever a governed file already merges through the driver, and names the line.

The prose skeleton folds blank runs like the roadmap's, and a section arriving in a block-organised file goes to the end of its own block.

Twenty-one merge scenarios ran against the driver, seven defects were fixed as their own tasks, and the strategy is in the skill.

The merge skeleton folds blank-line runs, so a block added on one side merges with lines filed under an existing one on the other.

SessionStart re-vendors a copy older than the engine install --vendor would pick and names both versions (design recorded in plugin/src/roadkeep/installing.py).

install --vendor keeps .roadkeep/reports through the engine swap, every capture byte-identical and counted (design recorded in plugin/src/roadkeep/installing.py).

A wholesale grammar failure names this engine and the newer wired one, with the restart; an empty [grammar] reads as declared (design recorded in plugin/src/roadkeep/linting.py).

The plugin ships a portable plugin.json and stdio mcp.json for Codex, stamped with one version; Codex reuses its hooks and marketplace, never twice (design recorded in tests/test_plugin.py).

install --host codex wires .codex/hooks.json, the config.toml server and .agents/skills, names the /hooks review; check and uninstall follow (design recorded in plugin/src/roadkeep/installing.py).

codex_skills renders the four commands as roadkeep-<verb> Codex skills with the real launcher and no Claude-only spelling; one main skill (design recorded in plugin/src/roadkeep/installing.py).

engines judges the enabled Codex copy beside the Claude plugin, names Codex's remedy and launcher, and a write pinned behind either is refused (design recorded in plugin/src/roadkeep/installing.py).

The MCP server binds to -C, then ROADKEEP_PROJECT, then cwd, says which on stderr, and refuses calls when started inside its own install (design recorded in plugin/src/roadkeep/serving.py).

Under Codex (turn_id or guard --host codex) a governed shell mention is denied, not asked; staging, history and roadkeep stay allowed (design recorded in plugin/src/roadkeep/guarding.py).

The guard decodes every apply_patch target, both ends of a move included, and denies a governed one; undecodable input gets the mention test (design recorded in plugin/src/roadkeep/guarding.py).

The gui site prerenders through prerenderSite 2026.3.26, its Markdown link via headTags, output byte-identical; both sites registered (design recorded in gui/site/scripts/prerender.mjs).

The site ships a checked llms.txt and JSON-LD, and shows askround's own dated round: named by no assistant yet.

site/docs builds the docs at /askround/docs/, its MCP and CLI reference generated from the package.

site.yml deploys the site to Pages; alegauss.github.io/askround/, its docs and llms.txt answer 200.

site/ prerenders the page at /askround/ with a real recorded round; Lighthouse scores 100.

The plugin, askround install and a vendored copy for the web wire every clone.

Every MCP tool's response is counted on the recorded rounds against a budget in budgets/mcp-tokens.json, and askround_run returns one brief per question.

A question asked again the same way within the day returns the stored round with its age and calls no provider; --fresh asks anew, and a round with failures is asked again.

With background: true a round or set starts as a job and returns at once with its estimate; askround_job reads its progress and result, and each answer is kept as it lands.

The person sets a day and month budget in a config no tool writes; a run past it is refused before any call as an approval request, and spending comes with every paid result.

Every pre-call check runs first and a refusal lists all problems with a code and its fix; askround explain covers each code, and MCP schemas carry the limits.

Every command's JSON names a versioned schema published in schema/ and validated in tests; each fact carries its answer id and passage range; exit codes are documented.

Agents get a brief: a round in a few hundred tokens with answer ids on every fact; askround_answer returns one answer or just the passage around a quote.

askround's six laws open IMPROVEMENTS.md as §0.1, from context is cost to the person sets the money, and the project skill says a rationale names the law it serves.

The suite runs with no key and fails any network call; adapters are tested on a recorded live round, and ASKROUND_LIVE=1 runs each keyed adapter against its API.

A skill tells an agent when to run askround itself (visibility, discovery, research), how to ask neutrally and within budget; a validated plugin installs it with the MCP server.

Research mode reports where the assistants disagree on a fact first, with both sides, their passages and the primary sources to settle it; one provider's own sources are not read.

askround research asks every provider with search, merges the answers' verified statements into claims ranked by provider support, with passages and sources primary first; also an MCP tool.

askround mcp serves ask, answer, run, report, diff and doctor over stdio with every refusal and the cost cap; a stdio client asked two providers and read an answer.

askround export writes a product's rows from the series per day, provider and question as CSV, JSON or a daily.json shape, and the README shows a weekly schedule on Windows and cron.

A run prints its estimate from a dated price table, refuses one above --max-cost before any call, and reports the actual cost per provider from each API's usage.

askround diff compares two rounds or set runs per question and provider, naming share and rank changes, rivals and cited pages that appeared or left, and claims made or dropped.

Every saved round appends a row per answer to series.jsonl, rebuildable from the rounds; askround series gives a product's share per week, day or month per provider.

Every question, seed, variant and lookup is scanned for emails, phones, CPFs, IPs, API keys and the person's private names, and stops before any call naming what it found.

A question naming the product, an alias, a place or a listed phrase is refused before any call; a brand question (--brand, brand = true) runs, marked and reported apart.

askround discover writes variants of a seed that never name the product, runs them and ranks them by rank and frequency; a 1-run live test ranked the seed first.

A product file's [[question]] entries form a prompt set; askround run asks them all as one set with a question-by-provider table, and examples/modelado.toml holds the four 12/11 questions.

A product file holds a product's facts and pages; with --extract the answers' claims are judged against them, and a contradicted or doubted fact is reported with the page that should say it.

--check-links follows each link and image an answer gives beside a product and flags those landing on another product's page or broken; Modelado's real Store, site and repo pass.

askround lookup and --check-rivals check a rival on GitHub, then a searched call; both 12/11 'invented' names exist (design superseded: they exist).

askround report turns a stored round (the latest by default) into one Markdown or JSON report: frequency and ranks per provider, claims, cited pages and rivals, without the answers' text.

With --extract each answer's claims about every product (table cells and prose) are read as attribute and value with a verified quote, and a product's claims are aggregated per provider.

Each answer's cited pages are normalised; given its domains, a product's cited pages are counted per provider by kind, and third-party domains and rivals' pages are counted too.

With --extract a round lists every rival the answers recommend, with answer counts, ranks, providers and a quote; on 12/11 Modly led, but a live round today ranks it lower.

--extract lists each answer's recommendations in order with quotes, checked against its text, so a product ranks by "Best fit" or prose; live and synthetic answers checked, 12/11 originals not kept.

Rate limits, server errors and timeouts are retried with backoff and Retry-After; a call that still fails is recorded with its attempts and the round finishes, listing failures.

--locale COUNTRY:LANG asks each provider as that locale through its search settings, records what its API ignored, and repeated locales are compared side by side.

Every answer records the requested and served model id, search tool, temperature, locale, settings, run and time; --model pins a model and frequency groups by served model.

A round asks each provider three times by default (--runs) and --product reports per provider the share of runs naming it and its position in each, raw answers kept.

askround doctor lists each provider's key variable (never its value), model, search tool and a free check call, telling rejected keys from missing models.

Every adapter searches with its provider's own tool and keeps queries and cited sources, with --no-search to compare; OpenAI and Gemini were checked live.

askround ask sends one question to every provider with a key (OpenAI, Gemini, Anthropic) in parallel, prints the answers together and saves each raw response with the date.

voice.lines lists matching sound.speak takes as adoptable, and adopt=true writes the line into their records so they count current (design recorded in docs/specs/world.md).

capture.web films a declared visit in Chromium as a WebM for video.cut, recording the URL and step marks, and refuses a broken page (design recorded in docs/specs/video.md).

An act whose application exits during the call or its read-back lands, saying the application exited, so the exit claim after it is what checks it (design recorded in README.md).

An in-app render of a window draws its client area at its client size, so no blank band runs down the right and bottom where its frame is (design recorded in README.md).

The file-arrives case writes from a thread of its own and gives the step 15 s, so a loaded guest still sees the file and the case reds only when it never comes.

A locator's order is read by name only, so [order=1], [order=9] and [order=-1] are refused with the four orders there are.

An in-app render of a backdrop window composes what shows there: the capture key, the theme's ApplicationBackgroundBrush, or the Mica base (design recorded in src/Winwright.InApp/Renders.cs).

A "focus" act puts the keyboard focus on its element through UI Automation and reads it back, so the press after it lands on that row (design recorded in README.md).

An act takes "passes": a key whose string must show after it, watched by UIA events and looks armed before it, so a flash is seen (design recorded in src/Winwright/Asserting/Passing.cs).

A red guest run brings back every dump: the test host's as hangdump.dmp, each other under its own name and process, and what a size cap left is named (design recorded in tools/blame-gather.ps1).

A process step "launched again" starts the app the way its fixture did after a claimed exit, without restaging files, so later steps read what the close kept (design recorded in README.md).

A press of a bare digit is a chord and reaches the window as that digit; the press reads traversal keys by name only, never by number.

A drag takes "holding" (ctrl, shift, alt, ctrl+alt), down from the press to after the release, with Alt's release masked by a Ctrl tap so no menu opens (design recorded in README.md).

A click step takes "at": "x,y" in drag's point grammar and presses there; a point past the element is refused naming its size, and nothing is pressed (design recorded in README.md).

A click step takes "clicks": 2 and double-clicks in one batch at one point; other counts are refused, and the fixture's strip opens a tab on a double click only (design recorded in README.md).

A miss names each shut door by its state ('not selected', 'collapsed') and lists only the doors under the step that resolved (design recorded in README.md).

A red run on a 'shell' or 'stale' desk no longer calls the tree fine: the note quotes vm-run.log's excused count, or says the runner cannot tell (design recorded in tools/desk-note.ps1).

The README's first screen leads with the site's token pitch and the measured cost, approved by Alexandre on 2026-11-06; the CI gate moved to its own section.

The readings case polls for a second, so even a loaded host reaches the seventh poll and the case reds only when a change of reading is lost.

The settles-late box commits 1500 ms after the write, so a loaded guest's act still reads what it wrote and the case reds only when the read-back is wrong.

The code is Winwright again: WW590, WW591 and WW597 are reverted with no paneproof alias, since no Paneproof release ever shipped, and renaming is now a non-goal.

A guest run's last lines, after the verdict, name the host path of the guest's log and of each file it brought back this run, so any tail of the console still points at the evidence.

init --questions lists every choice init takes with a default for the directory; a bare init on a terminal asks them and prints the equivalent command, and the skill reads the list.

init with no --prefix and no ids derives one from the origin remote's name, else the directory, says so with the override, and asks for --prefix when the name gives none.

init names every step it left unset with the verb that adds it, and --install, --merge and --gate run those verbs once the scaffold lands.

The skill ships filing.md, opened when a write targets a declared sibling: add -C into a checkout, or a shallow clone, rebase with the driver and a merge request, then the dep.

A ship names each sibling line it unblocked, and the sibling's session start names each ready line whose foreign dep shipped, offline from a checkout or the cache.

An add from another governed project writes (from: <project>) on the line; list --foreign lists those lines, and amend --reviewed drops the clause.

list --dated answers each line's proposing and shipping commit and date, and --history <bare clone> --at <rev> dates a git archive extract that has no history of its own.

A weekly schedule publishes the version main already carries when it has no tag, the v0 tag follows every release, and the README pins the Action at @v0.

A lock whose holder process is gone is taken on the next write and said so, on POSIX and Windows; a live holder is still refused, with how long it has held the lock.

A defer refused for length names the reason as the field to shrink, with the room budget --defer reports, and its foresee door is budget <id> --defer.

A <name>: <id> dep resolves against the sibling's ledger from its checkout or a shallow cached fetch of its git remote at ref, with a timeout, and lint never fetches.

block add refuses a label any governed file titles differently, writing nothing and offering the next free label as the retry; block list's door now carries the declared title.

Both sites draw their nav, hero, buttons, sponsor and footer from the design system, the gui landing its theme toggle, copy button and feature index too, with analytics after consent.

picture.upscale buys a larger picture from a Scenario upscaler, recorded against its source, and store.too-small names that call (design recorded in docs/specs/fetching.md).

picture.buy draws from a public Scenario model with no training, shaped by what the model declares and dry-run priced first (design recorded in docs/specs/fetching.md).

The item view takes Decide's drawn boxes, sent as revision.ask's marks, kept as a mask on the picture's digest, named in the first message (design recorded in docs/specs/acceptance-spec.md).

An open project takes the window's whole width, its list a share a person drags or moves by key, the item filling the rest with prose kept to reading length (design recorded in gui/README.md).

A picture fits the item panel beside its request box, says its real size, toggles to it in one click and opens alone at full size (design recorded in gui/README.md).

A project opens on its first inventory page, with a tab per kind counted by the payload, a path search and a list drawing only rows in view (design recorded in gui/README.md).

Each level field's help has a key from its path, and level.fields appends missing ones to the string table untranslated, naming changed ones (design recorded in docs/specs/engine.md).

A picture drawn at another shape than asked is named in the answer, in its record as mismatch.aspect_ratio, and in a batch's mismatches (design recorded in docs/specs/fetching.md).

A level set's most_rise bounds each level to a ratio of the one before; Spinhold's phases hold its owner's 1.5 there (design recorded in docs/specs/measurements.md).

picture.batch says per entry why it would be bought and the cost against the ceiling, buying nothing, then buys exactly what that plan named (design recorded in docs/specs/fetching.md).

A bought picture a person accepted keeps their verdict in its record, so picture.buy answers kept and spends nothing when its prompt is reworded (design recorded in docs/specs/fetching.md).

sound.loop records its track, and a borrowed file keeps its maker, original and step, which provenance.credits says as one chain (design recorded in docs/specs/provenance.md).

engine.preset_search now answers the sheet of lossy presets by its path, so a person is handed them to judge; a test holds it with no engine.

picture.buy sends a 4.0 ratio as the resolution of that shape, which 4.0 draws by, and refuses one it has no size of before spending.

reference.draft turns a transcription a person accepted into a draft level by the project's mapping, recorded and held to its schema (design recorded in docs/specs/fetching.md).

level.reader writes a typed GDScript reader from the level schema; Cottony's 20 levels and Spinhold's 3 phases read through it with no error (design recorded in docs/specs/engine.md).

Motion, samples and the warp-in gate count Resogun's phase 1 well enough for Spinhold, the owner judged; telling the ship's shots from pods is dropped.

Spinhold's phases pass level.validate, compile, probe and curve with no polyweave change, its probe reading waves by analysis: peak alive 39, 54, 75.

level.curve holds a level set to its declared shape, and search.sweep's level producer turns a level's keys, compiled and probed, toward its spec (design recorded in docs/specs/measurements.md).

A kind's warp keeps only tracks born at a flash just before, so Resogun's lull counts none (79 before) and its phase 282 tracks (design recorded in docs/specs/fetching.md).

Every Cottony family renders from its source mesh with dedust, settle, fill or its own cloth step inside the bake, and the exporters and prep.py are gone.

render.in_game draws a bake's model in the game at its rig beside the bake: a cone overlaps by 0.996, its light 1.07 of the bake's (design recorded in docs/specs/geometry.md).

level.compile holds a level source to its schema, runs the project's own command and records the file the game loads, or the source the game reads (design recorded in docs/specs/engine.md).