kody

Intent Intentional test-debt surgery: remove redundant test code across the repo without dropping use-case coverage. The goal is fewer, higher-signal tests that still pin every product behavior and contract: execute, search, billing/credits, packages, MCP surfaces, auth, migrations, and entitlements. Why The unit test tree had grown to about 314k lines, which slows reviews, , and every future refactor. The fat was not a clone farm. Exact duplication is only about 4%. Most of it is verbose filter matrices (one per row), copy-pasted fixtures, near-duplicate scenario tests, search-stack overlap between layers, and assertions on dead or unreachable code. Why this is about 70k lines, not "hundreds of thousands": the whole unit test tree was 313,868 lines. Cutting 200k+ would mean deleting most real behavior tests. After the first pass over all 21 buckets, a second cross-file pass over the largest remaining files found only 2–5% more to cut, which is good evidence that the redundancy ceiling is roughly where this PR lands. Summary Net: went from 313,868 lines in 1,204 files to 244,089 lines in 1,199 files, about −69,800 lines (−22.2%). Whole-repo diff is +76,106 / −145,854 (net −69,748). Test counts: node-unit went from 3,660 to 3,253 tests (same behaviors, fewer blocks), and workers-unit from 406 to 367. What was done, by technique: Table-driven rows: one-test-per-row filter and parse matrices became tables, asserted as so a failure still names the offending row. Merged workflows: separate tests that set up the same state and then checked one field each became one longer workflow test, per the house testing principles. Shared per-file fixtures: repeated 30–100 line inline env, row, and actor literals became small / builders in the same file. Expected values stay literal. Real SQLite with migrations instead of hand-rolled fake D1 where the fake re-implemented SQL (, , , , , identity tests). Deleted: tautologies, copy pins, tests of dead exports, duplicate worker and node doubles of the same assertion, per-test / boilerplate already covered by the global /, and blanket console silencing (replaced with narrow allowlists where needed). Five files merged away: and merged into . merged into . and merged into . Per-area commits (net lines), one commit per area: No production runtime changes. The only non-test files touched are three test-support helpers, each used only by the tests in its own area: , (which also fixes an undeclared reference), and . Billing, auth, isolation, and metering assertions were not weakened. Exact credit, rate, and count values stay exact. Cross-user isolation checks stay, and several are now stricter (exact , a check on every reset-client scenario, exact redirects). Referenced issues and PRs No issue drives this change. It is intentional test-debt surgery. Related to #2708 (repo-wide break, currently the only red check here) and #2709 (test files are not typechecked, filed from this work). Testing on the branch: format, lint, typecheck, knip, slop-ratchet, primitives, migrations, deploy-guardrails, workflows, docs, mermaid, lockfile, and all worker builds pass. test-node: 1,092 files and 3,253 tests pass. test-mcp passes. test-workers: 365 of 367 pass. The 2 failures were 20s timeouts under full-validate CPU load ( and ). Rerun alone, both files pass (8/8). e2e: 17 of 18 pass. timed out under load and passes when run alone. No e2e spec was edited. worker-startup-time exceeded its budget only under load. Alone it passes: platform 196 ms against a 340 ms budget, runtime 79 ms against 160 ms. Production bundles are unchanged. audit:prod fails on an existing moderate undici advisory that reaches transitively. It is unrelated: this PR doesn't touch . CI on : 🧪 Node, ☁️ Workers, 🔌 MCP, and 🎭 E2E are green. Preview deployed, and is ok at the PR merge SHA. 🧹 Static fails only at the production dependency audit step. Every other Static step passes (format, lint, typecheck, builds, startup, knip, slop-ratchet, and the rest). That audit failure is repo-wide: advisories published 2026-09-28 for undici, ip-address, and esbuild. Filed as #2708. Reviews: Cursor Bugbot found no issues. CodeRabbit skipped the review because 550 files exceeds its 150-file limit. Coverage, measured before and after with on node-unit (installed with ): Lines went from 67.82% (60,549 / 89,273) to 67.83% (60,569 / 89,295). Branches went from 58.77% to 58.78%. A per-line diff against the base tree found 9 source files that lost a few incidental lines or branches. I restored assertions for the meaningful ones (last commit): A provider-mark label-only save succeeds without logo storage and keeps the existing logo. Preview dry-run cleanup works with no Cloudflare credentials. The R2 bucket delete success path. Identity-icon refresh skips stamping for packages and returns early when bindings are missing. Residual risk Merged tests fail as a unit. A failure early in a merged workflow hides later scenarios in the same test until it is fixed. Table diffs and row labels still identify the failing case, but there are fewer test names. Test files are not typechecked by the repo gate. Every worker tsconfig excludes , so fixture and type drift in tests surfaces only at runtime. The agents checked changed files with a throwaway tsconfig and introduced no new errors, but pre-existing ones remain. This existed before the PR. Real SQLite fixtures now couple to the migration chain. A future migration that reshapes , , secrets, or flags tables fails those tests at seed time. That is an intended signal, but a noisier one. Shared per-file builders: changing a default in one builder affects several tests in that file. Defaults mirror the prior inline literals. Global mock config dependency: several files dropped explicit /reset helpers and rely on the Vitest config's + . If that config changes, state could bleed between tests. Some checks are stricter, for example exact spawned-call lists and exact request sequences. Harmless future behavior changes may need test updates. Coverage not restored, all incidental fallbacks or dead code: the dead export in , which has no production callers Thinner end-to-end coverage for webhook HTTP challenges: meta-hub and websub challenges are no longer exercised end to end over HTTP. Coverage relies on the GET path staying generic. System recap — composes existing primitives (low risk, test-only) Mode: recap · Base: @ · Head: Classification:** composes. No primitive is added or changed. The diff touches only files plus three test-support helpers (, , ). No production runtime file, migration, or lockfile changes. Primitives touched Change flow Each unit suite still exercises the same production primitives. This PR only changes how many test blocks and fixture lines it takes to do that. Invariants Per-user isolation, billing and metering, and auth assertions were kept or tightened, not loosened. Where fakes were replaced (fake D1 in , , , , , identity tests), the replacement is real SQLite with the real migration chain.   

Intent A 43.5s marketing motion piece for Kody that sells personal software, connected to everything, built with Remotion from the product art and tokens already in this repo. Why Every new personal app pays the integration tax again: sign-ins, pasted API keys, MCP configs. The video makes that pain concrete, then resolves it into Kody as the one home for connections, and shows that home turning out app after app that already has them. Summary New standalone project at . It is not an npm workspace and no worker imports it. Composition : 1920×1080, 30fps, H.264 High (bt709) with AAC, 43.5s. After the drop, one piece moves at a time. Beats: 0–10s, integration tax. "You need a new app." A checklist runs through one full cycle: "Wire up integrations" (Google sign-in, Stripe key, GitHub MCP, Slack app, and Linear key modals pile onto the app), then "Build" (the app fills in), then "Deploy" (it goes Live). Then "Now you need another one." and "Wire up integrations… again.", followed by three "and again…" lines, each faster, as more apps pile up with the same modals. It lands on "Ugh 😩" with every window slumping gray, just before the drop. 10–24s, connect once. The lantern lights on the drop. Then the six primitives light one per beat, each with its label. After the labels clear, the services fly into the lantern one at a time. Then "Connect once." lands, then "Keep them in Kody.", then the Secrets and Connections panels open. 24–32s, generate, then many. "Generate software that already has the connections." An agent prompt becomes the Morning HQ dashboard, fed by beams from the lantern. The camera then pulls back to 0.14× while 51 more apps pop in faster and faster around the lantern, and the reprised tallies read Apps ×52, New sign-ins ×0, API keys pasted ×0. The apps are 24 different kinds of interface (): chat, kanban, music player, route map, interval timer, booking, word game, home controls, photo album, invoice form, code review, chores, flashcards, video call, sketchpad, budget sliders, seat picker, poll, terminal, grocery steppers, editor, habit tracker, approvals, and journal. The first 24 to appear are all different, and repeats sit at least five cells from their twin. 32–40s, stays lit. The pull-back finishes and holds for a second, then "Agents come and go." replaces the headline. Next, Cursor, Claude, and ChatGPT each connect in turn on the right. Once the last one leaves, "Your software stays lit." appears, and only then do scheduled triggers wake batches of apps. 40–43.5s, close. "Personal software, connected to everything." plus the koala logo and kody.codes. The only outside visual asset is the 😩 image, a Noto Emoji SVG (Apache 2.0), which is committed so the render never depends on a system emoji font. Reuses the lantern still, the six orb sprites, the homepage orb layout and clip path (), the primitive words (), the dark-mode tokens, the Bricolage and Wix fonts, the logo, the brand pattern, and the in-repo service and agent marks. Sound has two layers, both keyed to the same cue frames as the visuals: is an edit of Kent's Suno track "Personal software" (D major), built by (, needs ffmpeg with librubberband). Its intro didn't fit the "ugh" opening, so the first 10s are sound design instead: a clock tick-tock that doubles its speed with the repeats, a low drone that creeps up, and a tape stop plus deflate on "Ugh". A reverse swell then leads the track in on its own riser, so its drop lands at 10s as the lantern lights. The pre-chorus and chorus play under the app build and zoom-out, and the track's final hit lands under the tagline at 40s. The edit points live in (). Each section is stretched from about 122 BPM onto the 120 BPM grid, and the mix is at -14 LUFS. places Kenney CC0 one-shots () on the cue frames: card pops that climb the D-major scale, window whooshes, checklist ticks on D, F#, and A, "Live", the "and again…" thuds, the "Ugh" slam, in-key pings as each primitive lights, pops as each service lands, a light whoosh per agent, and pings as triggers fire. The opening sits about 8 dB under the drop. holds Beat 1's cue sheet as plain data, so the scene, the effects, and the soundtrack builder all read the same frames. !Opening contact sheet !Beat 2 sequence: primitives, then integrations, then words, then panels !Beat 3: many kinds of app from one home !Beat 3 pull-back contact sheet !Beat 4 sequence: zoom settles, agent copy, agents one at a time, stays lit, triggers, tagline !Soundtrack waveform, one grid line per second The rendered master is in Dropbox at (direct link). Render output () and the downloaded Suno source () are gitignored. renders an MKV with PCM audio and has ffmpeg encode the AAC for the MP4, because Remotion's AAC output left the encoder delay in and played about 43 ms late. Referenced issues and PRs None. Testing in the project, plus the root , , , , , and . Rendered the full master and reviewed frames across every beat, including frame-by-frame through the pull-back and each headline handoff in the reworked opening. Rendered all 24 app screens side by side at close range to check their layouts, and checked that no two identical screens are adjacent in the grid. Audio: measured the final render at -14.3 LUFS integrated with a -1.5 dBFS true peak. Cross-correlated the MP4's audio against the soundtrack at 10s, 20s, 30s, and 40s: 0 ms offset (it was 42.7 ms before the MKV-plus-ffmpeg change). Kick onsets in the drop sit on the 0.5s grid with the same small lag the detector shows on the source. The level arc runs from -38 dB RMS at the start to about -17 at "Ugh", then the drop comes in at -13 and holds around -14 through the chorus. Checked the Dropbox copy against the local file: the SHA-256 of the direct-link download matches the local render. System changes System recap — composes existing primitives (low risk) Mode: recap · Base: @ · Head: Classification:** composes. No runtime primitive is added or changed. The new folder only reads static assets and pure TS modules from at bundle time. Change flow The Remotion bundle imports product art and homepage layout data from the worker package, then renders an MP4 offline.