Skip to content

fix: read cached source as utf-8 (fix #11682) - #11709

Merged
sheremet-va merged 1 commit into
vitest-dev:mainfrom
juan-fernandez:juan-fernandez/fix-cached-source-utf8
Oct 24, 2026
Merged

sheremet-va merged 1 commit into
vitest-dev:mainfrom
juan-fernandez:juan-fernandez/fix-cached-source-utf8

Conversation

@juan-fernandez

Copy link
Copy Markdown
Contributor

Prepared with AI assistance for @juan-fernandez, who reviewed the code and description.

Fixes #11682.

Read cached server modules as UTF-8 so Vite receives source text and stack formatting preserves the original error.

Regression coverage checks cached and uncached modules, UTF-8 text, and TypeScript source-map locations. Verified with Vite 8.0.11 and 8.3.5, including Node 24 and 26 reproductions.

Decode cached server modules as text for Vite source-map handling.

Prepared with AI assistance and reviewed by @juan-fernandez.

Fixes vitest-dev#11682
@netlify

netlify Bot commented Oct 23, 2026

Copy link
Copy Markdown

✅ Deploy Preview for vitest-dev ready!

Built without sensitive environment variables

Name Link
🔨 Latest commit 316d4e9
🔍 Latest deploy log https://app.netlify.com/projects/vitest-dev/deploys/6adb1799b3d371da227dfe85
😎 Deploy Preview https://deploy-preview-11709--vitest-dev.netlify.app
📱 Preview on mobile
Toggle QR Code...

QR Code

Use your smartphone camera to open QR code link.
🤖 Make changes Run an agent on this branch

To edit notification comments on pull requests, go to your Netlify project configuration.

@sheremet-va
sheremet-va merged commit 867a5e4 into vitest-dev:main Oct 24, 2026
31 checks passed
r-hashi01 added a commit to r-hashi01/astlide that referenced this pull request Oct 29, 2026
…rce (#286)

vitest-dev/vitest#11709 (for #11682): the server module runner read
cached module source as a Buffer; with Vite 8.3.5, mapping an error's
stack then crashed with "code.charCodeAt is not a function" and hid the
error. Applied with bun patch until a Vitest release includes it
(patches/README.md). The server image copies patches/ for bun install.

Refs #281

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Medformatik added a commit to OpenConditions/openconditions that referenced this pull request Oct 30, 2026
…#307)

## What changed and why

This is a round of dependency upgrades across the Node workspace, the
Python services, the Dockerfiles and GitHub Actions. Each moves to its
newest release that the repository's compatibility rules and
supply-chain policy admit.

The most important change is the runtime. OpenConditions still pinned
Node 24.21.0, which predates the October 2026 Node.js security releases
(Brotli decoder out-of-bounds write, TLS hostname verification bypasses,
keep-alive response confusion). `.nvmrc` and the three Node images now
use Node 24.22.2 (the Dockerfiles were already on 24.22.1), and pnpm
moves to 12.16.1. The openlr-resolver image moves to Python 3.14.10, the
expedited security release from 2026-10-25 that distrusts the
compromised eMudhra intermediate CA keys and hardens `tarfile`'s data
filter. CodeQL init and analyze move to 4.39.1, matching the upload
step.

In the workspace, Biome moves to 2.6, tsup to 9, croner to 11, and
Vitest, turbo, testcontainers, commitlint, eventsource and `@types/node`
to current patch releases. Transitive dependencies are refreshed within
their ranges. tsup 9 uses Chokidar 5, so the `trustPolicyExclude` entry
for Chokidar 4.0.3 is gone. Biome 2.6 formats Markdown and YAML by
default, so I turned that off: Prettier already owns those files, and
Biome's Markdown formatter rewrote frontmatter.

For the resolver, FastAPI goes to 0.162.0, pytest to 9.2.0 and Ruff to
0.17.3, in `requirements*.in` and `pyproject.toml` together. The hash
locks were regenerated from scratch with the reviewed compiler the
README prescribes: pip-tools 7.6.2 and pip 26.2.1 on Python 3.14.
Transitive changes are patch or minor releases: certifi 2026.10.23,
Starlette 1.10.0, Pydantic 2.14.2, websockets 17.3 and opentelemetry-api
1.45.2, plus dev-only tooling. FastAPI 0.162's one breaking change
concerns omitted boolean form fields, which the resolver doesn't use.
The TUF interop lock picks up pycparser 3.12.

Vite, which Vitest uses to load the tests, moves from 7.3.7 to 8.3.4. It
is held at 8.3.4 rather than the newest 8.3.5. Vite 8.3.5 changed how it
reads inline source maps (vitejs/vite#23775), and Vitest 5.0.5 hands its
`globalSetup` modules to Vite as a Buffer instead of a string. Any stack
trace taken inside the integration project's PostGIS global setup then
throws `code.charCodeAt is not a function`, and the run stops with "No
test files found". I bisected Vite 8.0.16 through 8.3.5, and only 8.3.5
fails. Vitest has merged the fix (vitest-dev/vitest#11709, upstream
issue #11682) but has not released it yet. The override comment says to
move to the latest 8.3.x once a Vitest release includes it. 8.3.4
contains the fixes for all five advisories the 7.3.7 pin was added for
(they are patched from 8.3.3). Vite 8.3 is the line that receives
regular patches, while 7.3 gets only backported important and security
fixes. The change adds Rolldown and its prebuilt platform binaries as
dev-only dependencies, with no install scripts. I measured no test-speed
difference: transforms are 1–2% of test time.

croner 11 changes DST handling, `startAt`, `dayOffset` and
`previousRun()`. Ingest only constructs jobs from cron expressions with
`catch: true`, so only the DST behaviour applies: a fixed-time job in a
spring-forward gap now runs at the first valid time, and interval jobs
run through a repeated hour.

Not upgraded:

- **TypeScript 7** stays excluded, as `dependabot.yml` documents.
- **`@openconditions/ocpi`** keeps its published `@types/node` on the
Node 24 line (24.19.2, the newest 24.x).
- The osmium toolchain (protozero 1.8.2, libosmium 2.23.1, osmium-tool
1.19.1) was already current.

## Verification and limits

- `pnpm verify:full` passed on Node 24.22.2, with Vite 7.3.7 and again
with Vite 8.3.4 on the identical lockfile: lint, typecheck, credential
and feed-schema generation checks, feed lint, 636 test files with 8,314
tests (unit and PostGIS integration; 17 existing skips), and the packed
OCPI consumer check.
- `pnpm audit:production --javascript` reports no vulnerabilities, and
`pnpm audit` reports none.
- Resolver CI steps, run in a Linux `python:3.14.10-slim` container:
  - dependency parity check;
  - hash-checked build and dev installs;
  - `pip check`;
  - Ruff over the resolver and image scripts;
- pytest, 67 passed (the four PostGIS reader tests ran with a
testcontainers host override because the run was
container-in-container);
- `pip_audit` on the build and runtime locks, no known vulnerabilities.

The TUF interop lock also installs cleanly. The script's Python audit
half needs `pip_audit` on the host, so I ran it in the container
instead.
- I built the openlr-resolver image locally on Docker Desktop (ARM). It
runs Python 3.14.10 with FastAPI 0.162.0, Starlette 1.10.0 and Pydantic
2.14.2.
- I also built the ingest image locally. It runs Node v24.22.2 and the
compiled osmium 1.19.1.
- Not run: hosted CI, image qualification and scans, the
contributions-api and federation image builds, and any deployment.

## Release and upgrade impact

There are no migrations or configuration changes. Deployments pick up
Node 24.22.2 and Python 3.14.10 with the next image build. Scheduled
ingest jobs follow croner 11's DST rules from then on.
nabobery added a commit to nabobery/fyrden that referenced this pull request Oct 30, 2026
With vite 8.3.5 and vitest 5.0.5, the API suite fails on Linux before
running a test: "No test files found", hiding an unhandled
"TypeError: code.charCodeAt is not a function" in vite's module runner.

vitest 5.0.5's server module runner reads a transformed module back from
its cache file without an encoding, so the module's code is a Buffer.
vite 8.3.5 started calling code.charCodeAt on that code while
source-mapping a stack trace. Any error whose stack passes through such a
module, here the chunked-import fixtures in globalSetup, then throws inside
Error.prepareStackTrace and ends the run. This is
vitest-dev/vitest#11682, fixed by
vitest-dev/vitest#11709 but not yet in a release (5.0.5 is the latest).

Reproduced in a node:24 Linux container from a clean checkout: with vite
8.3.5 every API run stops at "No test files found"; with vite 8.3.4, or
with 8.3.5 and the cached code read as text, the tests run. It did not
show on macOS.

vite stays on 8.3.4, which has no advisory, until a vitest release
carries the fix.

Signed-off-by: Avinash Changrani <74554800+nabobery@users.noreply.github.com>
omridevk added a commit to conciv-dev/conciv that referenced this pull request Oct 30, 2026
Bumps vitest, @vitest/browser-playwright and @vitest/coverage-v8 from
4.1.12 to 5.0.5. Patches vitest@5.0.5 with the unreleased upstream fix
vitest-dev/vitest#11709 (read cached server modules as utf-8), which
otherwise crashes error formatting with code.charCodeAt (#11682).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
omridevk added a commit to conciv-dev/conciv that referenced this pull request Oct 31, 2026
… on the library (#625)

* docs: plan for taking every dependency to latest

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* chore(deps): @tanstack/ai family to 1.0, harness SDKs to latest

Catalog moves every @tanstack/ai* package to 1.0 (ai, ai-client,
ai-code-mode, ai-persistence 1.0.1; the rest 1.0.0). minimumReleaseAgeExclude
collapses the per-package TanStack entries into one '@tanstack/*' pattern.
claude-agent-sdk 0.3.312, opencode sdk 1.19.4, modelcontextprotocol sdk
1.32.2. @agentclientprotocol/sdk stays on ~1.8.0, the range the bundled
@tanstack/ai-acp 1.0 is built against. The 0.x-keyed patchedDependencies
entries cannot apply to 1.0 and are dropped.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* refactor: call sites on the @tanstack/ai 1.0 API

TOOL_CALL_START drops the toolName alias (toolCallName only); a
TOOL_CALL_RESULT error is metadata.tanstack.state 'output-error' and success
carries no state. ApprovalRequestedEvent is gone, so the gate's
approval-requested chunk is typed as a plain CustomEvent. chat(),
useChat() and tool-bridge provisioning now require threadId; stories and
tests pass a stable one. A complete structured-output part carries data.
The queued-message helpers take the readonly queue the client now returns.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* chore(changeset): deps-ai

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* chore(deps): vitest 5.0.5 with local patch for vitest#11709

Bumps vitest, @vitest/browser-playwright and @vitest/coverage-v8 from
4.1.12 to 5.0.5. Patches vitest@5.0.5 with the unreleased upstream fix
vitest-dev/vitest#11709 (read cached server modules as utf-8), which
otherwise crashes error formatting with code.charCodeAt (#11682).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* chore(deps): tanstack start 1.169 set and site/example deps to latest

TanStack router/start set to latest (react-start 1.169.4, react-router 1.170.46,
router-cli 1.167.46, router-plugin 1.168.48, solid-router 1.170.43, react-form 1.33.5,
redact 0.2.3). Start 1.169 removed the built-in useSession; the try pairing token and
dismissal flag move to Start's getCookie/setCookie helpers (token zod-validated as a uuid).
redact 0.2's vite plugin now owns the react-dom/server and use-sync-external-store aliases,
so the e2e app keeps only the static entries it does not cover.

Site: fumadocs-core/ui 16.17.8, fumadocs-mdx 15.4.10, radix-ui 1.8.1, shadcn 4.23.0,
tailwind-merge 3.7.0, takumi-js 2.18.2, wrangler 4.154.0, @cloudflare/vite-plugin 1.65.0,
lenis 1.3.27, geist 5.3.0, magic-move 4.5.0, mermaid 12.2.0 (also ui-kit-chat).
@huggingface/transformers held at 4.2.x: every 4.3 pulls onnxruntime-web 1.31 whose
asyncify wasm is 25.6 MiB, over the Workers 25 MiB asset limit, which fails prerender.
nextjs example: eslint 10.13.0, eslint-config-next 16.4.3.

mermaid 12 bundles elkjs; the extension-testkit host's UnoCSS pipeline scanned it from
node_modules and emitted an unparsable selector, so node_modules is excluded there.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* chore(e2e): ignore the AGENTS.md next 16.4 dev writes

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* chore(deps): e2e consumer apps and whiteboard extension to latest

e2e/solid-start: @solidjs/start 2.1.0, @solidjs/router 1.0.0, @solidjs/vite-plugin-nitro-2 0.3.2;
start 2.1 peers vite ^8 so the app now takes vite from the catalog.
e2e/svelte: @sveltejs/kit 3.1.0, adapter-auto 8.0.0, vite-plugin-svelte 7.3.1, svelte 5.57.3,
svelte-check 4.7.7. e2e/astro: astro 7.4.2 (engines follow astro's >=22.23.4).
whiteboard: @excalidraw/mermaid-to-excalidraw 3.0.0, @takumi-rs/core + helpers 2.18.2,
@libsql/client 0.18.0, cytoscape 3.34.4, drizzle-orm 0.45.4, drizzle-kit 0.31.11.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(e2e): kit 3 migration, astro dev foreground, drop the solid-start vite exemption

SvelteKit 3 removed $lib: the app declares #lib in package.json imports and its tsconfig
extends $app/tsconfig. astro 7.4 auto-backgrounds astro dev when it detects an AI agent,
so the playwright webServer exited early; --ignore-lock keeps it a foreground one-off
server. solid-start no longer needs its own vite line, so check-catalog drops that
exemption and its test exercises the remaining whiteboard drizzle exemption.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* chore(deps): root tooling, changesets 3, fallow 3.36, publish deps to latest

turbo 2.11.10, @j178/prek 0.5.7, @arethetypeswrong/cli 0.18.5, fallow 3.36.0 (action pin
moved to the v3.36.0 commit), @changesets/cli 3.0.3, @changesets/changelog-github 1.0.1,
@changesets/parse 1.0.0, execa 10.2.1, js-yaml 5.5.0 (ships its own types, @types/js-yaml
dropped), minimatch 10.2.6, @ast-grep/napi 0.51.0, nypm 0.6.12, @tanstack/intent 0.5.6.

changesets 3: privatePackages.version defaults to false now, set it back to true so private
@conciv/* stay in the fixed group; 'changeset tag' is 'git-tag'; publish passes
--no-git-checks itself and rejects unknown flags; changesets/action v2 inputs renamed
(version-script, publish-script, commit-message, pr-title). intent 0.5 requires reference
files to be linked from SKILL.md.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* chore(deps): build plugins to latest (babel-preset-solid, dce, magic-string, unplugin, chokidar)

babel-dead-code-elimination 2.0.0, babel-preset-solid 1.9.16, magic-string 1.4.5,
unplugin-solid 2.0.0, chokidar 6.0.0, unplugin 3.6.0, launch-editor 2.14.2.

Held: @babel/core and @babel/preset-typescript stay 7.29.7 (babel-dead-code-elimination 2
and babel-preset-solid still build on babel 7). unplugin-solid stays on 2.0.0 (3.0 needs
Solid 2).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* chore(deps): UI libraries to latest (@pierre/diffs 1.6, hotkeys 0.18, rrweb 2.2, pretext 1.0)

@pierre/diffs 1.6.2: the third File/FileDiff constructor argument is now the owner render
callback instead of a container-managed boolean; the Solid wrappers pass a callback that
re-renders with the current props. @tanstack/solid-pacer 0.24.0. @tanstack/hotkeys 0.18.0 and
@tanstack/solid-hotkeys 0.17.0: RawHotkey is RawHotkeyChord and a parsed chord is already a
registerable chord. rrweb, @rrweb/types, @rrweb/rrweb-plugin-console-record, rrweb-player 2.2.0,
js-beautify 2.0.3, @chenglou/pretext 1.0.0, fonts, @zag-js/json-tree-utils and focus-visible
1.47.0, @floating-ui/dom 1.8.0, @chromatic-com/storybook 5.6.0, @testing-library/dom 10.4.2.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* chore(deps): bippy 0.7.3 and react-grab 0.2.0

bippy 0.7 dropped getFiberStack, getNearestHostFiber, getTimings and getFallbackOwnerStack.
The react bridge walks the return chain itself, finds host and composite fibers with
traverseFiber (descending then ascending, as getNearestHostFiber did), and uses the renamed
getFallbackParentStack. The render tracker computes self time from actualDuration directly.
Lockfile for the whole lane.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(solid-diffs): route @pierre/diffs owner render requests through syncRender

syncRender returns the owner render callback the 1.6 File/FileDiff constructors take, so a
rerender the instance asks for renders the current payload from the one place that already
owns rendering, instead of an untracked props read inside each wrapper.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(ui-kit-chat): resolve turn metrics only from a connected viewport

@chenglou/pretext 1.0 rejects a font string that is not a canvas font. A detached viewport
has no computed style, so the probe read an empty font and pretext threw; 0.0.8 measured that
with the canvas default instead. The estimator now returns no estimate until the viewport is in
the document, so metrics always come from real styles.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(deps): re-patch @tanstack/ai-sandbox and ai-sandbox-local-process on 1.0

1.0 still declares the projection capability on every withSandbox run while
providing it only for a workspace, so each workspace-less turn ended in a
RUN_ERROR from the engine's provides check, and still builds the projection
root through resolveHarnessCwd. The local-process provider still forgets its
handles, so destroy(id) cannot kill the CLI it spawned. Both 0.x patches are
ported to the 1.0 dist; the stdin error-listener hunk is dropped because 1.0
ships it (src/handle.ts:918).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(recorder): build the replay player once its container is connected

rrweb 2.2 creates the replay iframe sandboxed and throws when the root is not in a document.
The saved-recording player mounted from a ref callback, before the dialog attached it, so the
replay never rendered. Both players now build through one whenConnected step, and the style
scope is read once connected so rrweb's injected styles are found in the right root. The live
tick feeds events through one helper and scrub steps come from a lookup, which keeps both
callbacks under the complexity gate.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* docs: spec for approvals on the @tanstack/ai 1.0 deny-and-retry flow

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* test: migrate browser assertions and configs to vitest 5

toHaveTextContent is now strict equality: partial and RegExp checks move
to toMatchTextContent, or assert the full text where it is stable.
Locators match exactly by default: substring getByText calls pass
exact: false. packages/extensions/page sets its React JSX transform via
oxc. Failure screenshots and attachments now land in .vitest/.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* chore: changeset for the vitest 5 upgrade

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* chore(deps): dedupe the merged lockfile

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* docs: @conciv/ai package spec; ground the approvals spec in the TanStack AI docs

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* docs: record @conciv/ai decisions

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* test: exact-match locators and text for vitest 5 browser tests

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* test: exact grab source lines and real absence checks for the page-session card

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* test(core): a late stop must not cancel the turn claimed after it

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* feat(ai): @conciv/ai package with messages, events, tools, usage and testkit subpaths

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(core): chat.stop cancels only the runs the client saw

A stop RPC and the next turn travel on different connections; a stop that
landed after the next turn was claimed cancelled every active run of the
session, including the new one. chat.stop now takes runIds, the server stops
only listed runs active in that session, and the client's flush awaits the
stop before sending. sessions.delete keeps the session-wide stop.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* test(core): a stop must cancel a run whose viewer already left

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(core): keep the cancel watch alive after the viewer leaves

runStream stopped its cancel poll in the consumer's finally, so a viewer
disconnecting (the client's local stop) killed the only in-process path that
turns a recorded cancel into an abort. The stop RPC then waited out the
15s settle timeout while the run kept going. The producer already stops
the watch when the run ends.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(ai): message and part types alias the ai-client UI shapes

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* refactor(protocol): approval, run lifecycle and usage move to @conciv/ai

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* refactor(ui-kit-chat): types, events and story connection from @conciv/ai

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* refactor(extensions): message types and defineTool from @conciv/ai

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* refactor(app): message types, events and testkit from @conciv/ai

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* refactor(harness-testkit): event types and chunk builders from @conciv/ai

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* refactor(core): chat types and conciv events from @conciv/ai

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* chore: changeset covers the detached-run cancel fix

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* chore(deps): lockfile for the @conciv/ai manifests

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(test): prebundle @tanstack/ai-client behind the solid-excluded @tanstack/ai-solid

@tanstack/ai-solid 1.0 ships a `solid` export condition (./ui), so
vite-plugin-solid's vitefu crawl adds it to optimizeDeps.exclude. Its
imports (@tanstack/ai-client -> @tanstack/ai/client -> partial-json) are
then served raw, and partial-json is CommonJS, so the browser fails with
"does not provide an export named 'parse'". Include the excluded package's
deps with vite's nested `a > b` specifier so the whole chain is prebundled
with CJS interop, as it was on 0.51.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* feat(ai): @conciv/ai/adapters with the CLI text adapters

TextAdapter, makeTextAdapter (DelegatingTextAdapter moves with it) and one
typed factory per CLI (claudeCode, codex, opencode, acp) with their option,
permission and model-list exports. @conciv/ai now bundles the @tanstack/ai-*
CLI adapters (noExternal, as harness did) and keeps @tanstack/ai and
@tanstack/ai-sandbox external; the #107 dist guard moves with them. Types build
through tsdown so the bundled adapter types inline instead of pointing at
devDependencies. testkit gains runAdapterTurn.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* refactor(harness,protocol): build on @conciv/ai/adapters

Harness keeps its per-CLI product knowledge and only swaps imports to
@conciv/ai/adapters, events, messages and testkit; protocol's HarnessChatConfig
adapter is TextAdapter. Neither manifest lists @tanstack/ai* any more, and
makeTextAdapter call sites in core tests and harness-testkit import it from
@conciv/ai/adapters.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* docs: layer @conciv/ai as primitives under a new @conciv/chat package

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(deps): @conciv/ai as a runtime dependency where dist reaches it

The embed bundle keeps every @conciv/* external, and its mount-impl
chunk imports @conciv/ai/tools and @conciv/ai/usage, so embed must
depend on @conciv/ai or consumers (extension-tanstack, dev hosts)
cannot resolve it. extension-page (./client) and ui-kit-chat-tools (.)
publish declarations that import @conciv/ai types, so it moves from
devDependencies to dependencies there.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(storybook): prebundle the ai-client chain through ai-solid like the vitest hosts

Storybook now uses TANSTACK_AI_SOLID_DEPS, so @tanstack/ai-solid is
served through its solid condition and its ai-client entries (root,
byok, devtools) are prebundled, matching apps/conciv and ui-kit-chat.
lucide-solid resolves to its solid .jsx source, which vite cannot
prebundle ("Cannot optimize dependency"), so the include was a no-op
in both storybook and browserOptimizeDeps; drop it.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* feat(ai): @conciv/ai/server primitives and the node server-testkit

The turn runner (chat + withSandbox + caller middleware), the tool gate
middleware, WebSocket/SSE turn delivery with resume, run control (run store
types, cancel, memory durability, keyed locks, an async queue), the live
message fold, code mode and its isolate drivers, the local sandbox and the
command policy, typed with @conciv/ai primitives only. @conciv/ai now bundles
the patched @tanstack/ai-sandbox-local-process.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* refactor(chat): move the chat plane out of core and db into @conciv/chat

@conciv/chat holds turns, hydrate, transcript import, stop, compaction, boot
recovery, the permission gate, the session bus and the chat stores over
@conciv/db's tables (run store, thread queries, message and metadata stores),
every @tanstack/ai call going through @conciv/ai/server. Core wires
createChatPlane and passes the read-only command policy in; the MCP and push
surfaces stay in core. db keeps the schema, migrations and the legacy import
run at open. core and db drop @tanstack/ai*. Behaviour unchanged.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* chore(release): publish @conciv/chat and extend the @conciv/ai changeset

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* docs(spec): one entry per package, internals in concern folders

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* docs(spec): chat plane on the library's run

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* docs(spec): record the user's chat plane decisions

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* docs(spec): @conciv/ai/server is one abstraction

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* docs(spec): tests through public entries only

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* refactor(ai): three entries (., ./server, ./testkit) with internals in folders

Merges the messages, events, tools and usage subpaths into the root entry and
adapters into ./server; every import site moves to the new entries.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* docs(spec): child-call id patch, MCP discover_tools, release-age trim

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* docs(spec): approval tools stay out of code mode

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* docs(spec): decisions after the second build

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* docs(spec): ai owns the chat tables and stores

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* docs(spec): decisions after the third build

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* docs: format the deps-latest plan and specs

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* feat(chat): chat plane on the library's run

chat() with withPersistence, withLocks and withSandbox plus one conciv
middleware; library delivery, requestRunCancel stop, reconstructChat
hydrate, discover_tools. Approvals are deny-and-retry and approval tools
stay out of code mode. The 82 chat test files become chat.test.ts and
chat.it.test.ts.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* test(extensions): move extension tests off external_catalog

Tests call external_<name> directly and use discover_tools, reached
through a new makeDiscoverTools testkit helper. Drops the two push-channel
tests that pinned out-of-band approval asks, which the approvals spec
deleted.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* test: fold the lane's chat and ai tests into existing files

chat.it goes to core's rpc/wire.it, the chat plane unit tests to core's
app-harness-di, and packages/ai gets a single ai.test.ts. @conciv/chat
no longer has a test script.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* feat(ai): chat session on @conciv/ai/client, boundary lint rule

createChatSession moves from packages/client to the @conciv/ai/client
entry, unchanged, with its rpc and connection handed in. createChat
replaces useChat in stories and tests. Every @tanstack/ai* dependency
outside packages/ai is gone and oxlint now forbids the imports. The
test MCP client moves to harness-testkit on the MCP SDK.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(harness,try): agent prompts name discover_tools, not external_catalog

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* chore(deps): trim minimumReleaseAgeExclude, pin nitro-nightly past the window

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* chore(deps): oxlint 1.90, oxfmt 0.75, eslint plugins into the catalog, and reformat

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix: new oxlint 1.90 hits in React code, solid reactivity, empty export, toSorted

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* feat(chat): discover_tools shows each tool's output shape and declared errors

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* lint: scope React compiler rules to React code, restore LogoLoop watchImages

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* docs(skills): point debug and harness skills at the library chat plane

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* build(ai): bundle the agent sdks and the sandbox into the server entry

The CLI closure reaches @conciv/ai through @conciv/protocol, and the prod-closure guard
bans agent sdks. server.js is the only entry importing them, so they move to devDependencies
and inline like the other adapter packages.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* test(protocol): delete the empty usage types describe

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* test: fix the nextjs init clone and the tanstack resolution fixture

Next's AGENTS.md ignore moves to the root gitignore so the init clone does not inherit it.
The resolution fixture declares isolated-vm in allowBuilds because pnpm 11 fails the install
on an ignored build.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* test(e2e): refresh the nextjs packed fixture lock

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(example): tanstack-start is React again under the framework override

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(page): pin bippy to 0.6.1, one copy with react-grab

0.7's main entry imports react, which breaks the embed in hosts that do not
prebundle it. The 0.7 call-site changes are reverted. bippy/source 0.6.1 has no
sync fallback stack, so rawFrames falls back to the public async getParentStack.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(chat): tools in code mode read their tool call id from the library context

The approved one-line patch hands the minted callId to the binding, chatTool
puts it on the ToolRequest, so captures store under the id on the wire again.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* refactor: protocol is the bottom layer again

The library-derived chat, event, tool and usage types move from @conciv/ai to
@conciv/protocol, which depends only on @tanstack/ai and zod. The harness
contract that needs TextAdapter moves to @conciv/harness. @conciv/ai loses its
root entry, keeps the agent sdks and ai-sandbox as normal dependencies, and
makes ai-isolate-node optional behind the QuickJS fallback.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* test: approval asks assert the pending-approvals strip

The signed approvals spec puts asks in the strip, not the tool card. The calm
test now waits on the strip, approves, and checks the retry leaves every other
surface in place; the kit test checks the ask stays out of the trace.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(client): a rejoined run reports its lifecycle

joinRun bypassed the noticing wrapper, so a pane that joined a run another
client started never saw the terminal lifecycle and narrated forever.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix: settle the layering move

Protocol self-imports stay relative, @conciv/ai tests read their own entries,
and tool part metadata is narrowed now that it is library-typed unknown.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* refactor: keep @conciv/ai out of the CLI closure

ui-kit-chat only needed client types at runtime-free positions, so SendContent
and QueuedMessage move to protocol and @conciv/ai becomes its dev dependency.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(ai): a refresh keeps the ids of the rows on screen

Hydration hands assistant messages new ids, so stopping a run remounted the
trace. The refreshed transcript keeps each rendered message id by position.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* test: the approval pause holds still and its retry keeps the turn rows

A retry is a new turn, so the previous trace folds by design. The calm check
covers the pause itself and the retry keeps both turn rows mounted.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* test(embed): count messages inside the chat log

The session title now derives from the first message, so the bare text
locator also matched the header, the selector trigger and its option.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* chore(protocol): drop the unused TtyCommand import

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* feat(core): refuse non-widget callers of approval tools, deny-and-retry

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(ai): bundle the patched ai-sandbox and ai-code-mode so consumers get the patches

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(ui-kit-chat): depend on @conciv/ai so published types resolve

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* chore: drop the jazz-tools protobufjs and esbuild overrides, correct the @conciv/ai changeset

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(core): count session messages from the chat thread; restore the calm approval pause checks

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(ui-kit-chat): type the chat binding from protocol so published types resolve

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* docs(skills): point the harness skill at the moved contract

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* test(app): prebundle gsap so a late mascot import cannot reload the browser suite

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* test(app): approval replay on the deny-and-retry flow

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(chat): outside approvals retry without a turn, settle always resolves, delivery split

An ask recorded outside a chat turn stores a one-time grant only that caller's retry consumes, and
permissionDecision tells the widget whether to post the retry message. MCP asks share admitCall.
settle survives a throwing end hook. Socket and SSE delivery move to delivery.ts. Run lifecycle
reaches open panes over the push channel, and an adopted native session keeps its derived title.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* docs(skills): debug skill follows the delivery split and outside approvals

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* chore(e2e): refresh the nextjs packed fixture lockfile

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* test(e2e): init clones copy only tracked files

Other e2e runs leave AGENTS.md and .conciv inside the app directories, so the clone
inherited them depending on task order. Copying git-tracked files replaces the skip list.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(plugin): compile instrumentation in the first webpack dev build

next dev --webpack only compiles instrumentation.ts when its watcher has
already reported the file by the time the server prepares. A scan that lags
by about 80ms leaves instrumentation.js unemitted, Next memoizes the miss,
and register() never runs for the process, so the engine never starts.
Emit the entry in the first server build, which Next awaits before prepare.

The packed spec gains a cell that delays the watcher's stat of
instrumentation.ts, and the engine timeout now reports whether
.next/dev/server/instrumentation.js exists.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(chat): run events travel on the run log only

The server no longer broadcasts run lifecycle over the push channel. When a
run starts it publishes one session-level conciv.session-run {runId} event.
A client that did not start that run re-hydrates the transcript and tails the
run through joinRun on its live subscription, so content, lifecycle and
errors all come from the run log.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(chat): hydrate waits for a settling run before importing the CLI transcript

Stop then refresh read the transcript between the run turning aborted and
its end hook advancing the import anchor, so the stopped prompt was imported
again as a second user message. A run now stays live until its end hook
finishes, and hydrate waits for settling runs before syncing.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(ui-kit-chat): ChatProvider follows a swapped chat so queue Remove works after a session switch

The panel route passes chat() into one long-lived ChatProvider, which captured
the first chat it saw. After New session the queue strip showed the new chat's
queue but Remove cancelled on the old chat, so the item still ran. The provider
now keys its scope on the chat it is given.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(ui-kit-chat): a sent image shows in the user turn

The user turn mapped only document parts to a card, so an image the user
sent rendered nothing in the thread. Images now take the same attachment
card, which falls back to the image tile.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(chat): the compact marker counts the messages the widget shows

compact stored the raw thread length, but the widget gets the thread with a
run's assistant segments joined, so after any tool turn the marker pointed
past the end and the compact divider never drew.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* test(core): table the fake CLI's step tails so the adapter stays under the complexity gate

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(ai): Stop saves the text the client already saw

withPersistence throttles streaming snapshots to about once a second and does
not write the remainder on abort, so a stopped turn rehydrated short. The
conciv-turn middleware follows the streamed assistant text and, on abort,
rewrites that message in the library's message store. Stop now returns once
the run has settled, so the hydrate that follows reads the flushed text.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(ui-kit-chat): mermaid diagrams take the widget's colors

Mermaid rendered in its own lavender default. The diagram now uses mermaid's
base theme with themeVariables resolved from the --chat-* tokens at render
time, so light and dark both follow the widget scheme. No new tokens.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(conciv): the approval strip names what a CLI ask is for

A CLI ask carries only provider and toolName beside its title, so the strip's
first-string-argument target read "claude-code". The strip now renders the
ask's title.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(conciv): the slash picker ranks by name and Enter takes the top match

matchingItems sorted only by group, and the listbox kept a stale highlight
while the query narrowed, so /context + Enter ran /21st-ui-build. Matches now
rank exact name, name prefix, name contains, then description, and the
listbox highlights the first row whenever its collection changes.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(conciv): the status bar reads RUNNING for a server-driven run before its first chunk

The bar keyed busy off the stream's RUN_STARTED (sessionGenerating), which a
CLI adapter emits only once the CLI prints; the thread already keyed off the
server run lifecycle. The bar now also counts sessionRunning, so a compact run
shows RUNNING from start to end. The scripted harness gains a start gate to
hold a run before RUN_STARTED.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* test(core): table the fake CLI's step bodies so the adapter stays under the complexity gate

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(ai): Stop returns at once; the hydrate after it waits for the stopped run's flush

Stop no longer awaits the run settling. The thread read in hydrate waits for any
run of that thread whose controller is already aborted, so the text flushed on
abort is what the client reads back, with no sleep or poll.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(conciv): Refresh stays off for a server-driven run before its first chunk

The session names "a run is in progress, local or server" once, as
runInProgress, and the status bar, the refresh button and the session-menu
Refresh row all read it, so none of them opens a gap before RUN_STARTED.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* feat(conciv): a stopped turn says so in the thread and the status bar

The session derives stopped from the run lifecycle (phase aborted), so it
survives a reload and is never a stored message. The thread ends a stopped turn
with a quiet Stopped divider, reusing the divider styling, and the status bar
reads STOPPED instead of DONE.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(whiteboard): a canvas draw with no open canvas fails with NO_CANVAS instead of a pending id that never lands

Draws convert in the browser canvas island, so a draw now waits for the island to drain
its pending row and, past the deadline, removes the row and tells the agent to ask the user
to open the canvas.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* feat(chat): hand command permissions to the CLI's native mode and delete conciv's command policy

Claude Code runs with permissionMode auto (through modelOptions, the adapter's first choice).
What it escalates still reaches the approval strip through --permission-prompt-tool and the
default ask policy. Conciv's own approval tools keep the deny-and-retry flow.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(ai): images leave code mode as image blocks and page_screenshot saves a project file

The library tool bridge stringifies every result, so a custom provisioner (its
ToolBridgeProvisioner capability) lifts nested image parts into MCP image content. Image
bytes no longer count against the text cap, so child cards keep them through reload.
page_screenshot also writes .conciv/tmp/screenshots/<id>.webp and returns the path.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* docs,test: correct permission docs and tidy the wire and whiteboard tests

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix: carried highlight, stale queue flush after a later Stop, fallback fiber root

The trigger menu rebuilt its collection on every dispatch, so highlightFirst
pulled the highlight off carried items while loading. The collection now
only changes when the items do.

interruptAndFlush only sends if its Stop is still the latest one, checked after
the wait and before each queued send.

scannedRootFibers returns the current HostRoot, not the child below it, so every
top-level branch stays reachable.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* test(try-it): the health server teardown drops open browser sockets

server.close waits for sockets that never sent a request, which Node does not
count as idle, so an attached browser held the run open forever.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* test(conciv): queue flush helpers read last in the queue strip test

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(conciv): a superseded queued flush returns to the composer draft

interruptAndFlush resolves with the items it did not send, and the pane appends
their text under the draft and re-adds data images as attachments. Nothing
auto-sends and typed text is kept.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* test(ui-kit-chat): now-line asserts any spinner glyph, not one frame

The spinner cycles every 90ms, so waiting for one frame can miss it. The tests
now accept any braille frame, and cycling asserts the glyph changes.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* test(app): prebundle the shell hotkeys and mutation-observer deps

A late optimize of these reloaded the browser suite mid-run and left the next
test iframe never ready.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(ai): a socket turn keeps running when its page reloads

toWebSocketStream ties the producer to the socket: a socket close aborted
chat(), withSandbox detached the run (detachOnDisconnect), and with no
takeover driver or reaper the row stayed running with an open log, so the
reloaded page tailed it forever and Stop could not reach it. The socket turn
now drives the run the way the SSE transport does (its own durable drain,
decoupled from the viewer) and the socket only tails the run's log from the
start, so a reload rejoins a live run and the run finalises its row and text.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(ai): a stopped turn keeps its tool calls in the thread

withPersistence saves nothing on abort except tagging throttled text
snapshots, so a turn stopped mid tool call was never stored and the next
turn's thread dropped it. The run is folded with the library StreamProcessor
and on abort its assistant messages are stored with stopReason aborted,
replacing any snapshot copy.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* feat(harness): claude asks before risky commands through its native ask rules

Auto mode ran force pushes, recursive deletes and remote branch deletes
silently. The claude harness now writes Claude Code permissions.ask rules to
a settings file and passes it with --settings, so those commands reach the
approval strip. Verified live: force push asks, approve completes, deny
blocks, git status runs.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(conciv): every stopped turn keeps its Stopped line in the thread

The divider only followed the last run's lifecycle, so the next answered turn
erased it. It now also reads the library's stopReason aborted metadata that a
stopped turn is saved with and hydrate already carries, under every stopped
turn in history; the live turn still shows it the moment Stop lands, and the
trailing divider is the same one, so it never shows twice.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(chat): Deny re-runs the turn with the refusal, like Approve

The library flow re-runs chat() with the decision for both outcomes. Deny
on an in-turn ask now records a one-run denial, threads it to the run as
approvals false, and posts "Denied: <title>", so the agent learns it was
refused. Conciv's own ask tools fail with a final denial instead of asking
again. Asks from outside a turn keep no message and no turn.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(harness): ask rules also catch git -C and git -c forms

Claude Code matches rules against the command text, so git -C dir push
--force skipped every git rule. Each git risk now has a git * form too.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(ui-kit-chat): a thread row created detached is measured once it attaches

measureRow skipped any row that was not connected yet, so rows a pending
suspense boundary renders offscreen (a remount: the turn prefix reads the
markers query) were never handed to the virtualizer and kept their estimate
forever; the next row was placed over them and the trace header sat on the
prompt. virtual-core already observes a detached row and measures it when it
connects without caching 0, so the guard only hid rows from it.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* test(conciv): stop the turn only once its reply has streamed

The stopped-line case pressed Stop as soon as the button showed, so whether the
stopped turn had saved any reply depended on timing; a turn stopped before any
output has no stored message to carry the stopped mark.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(ai): decision messages drop the ask title's Approve prefix

The claude adapter titles asks "Approve Bash: <command>" and the strip shows
that title, so the posted message now strips the verb: "Approved: Bash: ..."
and "Denied: Bash: ...".

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(chat): clearing a session deletes only the metadata keys it owns

clearThread matched keys with LIKE '<threadId>:%', where _ in a session id is a
wildcard and the match ignores case, so clearing conciv_a_b also deleted the
grants and pending approvals of conciv_axb and conciv_A_B. It now compares the
literal, case-sensitive key prefix.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* build: require node >=22.23.4, which astro 7.4 needs

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(ai): a stopped turn saves only the tool calls the thread lacks

saveStoppedTurn stored the folded run whole, but the sandbox tool history had
already saved its finished tool calls as their own message, so after Stop the
hydrated turn carried every finished call twice and the cancelled turn grew a
second trace. It now saves only the parts the thread does not hold yet: the
in-flight call and the streamed text.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* test(conciv): the failed-chip send clicks the send button by its full name

Vitest 5 locators match an accessible name exactly, so 'Send' no longer finds
the 'Send message' button that main's test was written against under vitest 4.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* docs(skills): conciv-debug cites the current delivery and chat plane lines

The resume citation pointed past the end of delivery.ts. It now cites resume,
the socket tail and the plane's hydrate and routes at their current lines, and
says a socket turn drains into its log on its own so a closed socket never
stops it.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* build: published packages keep their node >=22.13 floor

Only the workspace root and e2e/astro need node >=22.23.4 for astro 7.4; the
published packages and e2e/solid-start go back to their own floors.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* ci: move workflows from Node 22 to Node 24

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* ci: carry the Node version in turbo cache keys

Turbo does not hash the Node version, so a Node change would replay old results.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(ai): a driven run seals its log only after turn-end writes land

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* test(recorder): wait for the control subscription before signalling it

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(ui-kit-chat): a follow landing no longer erases the reader's scroll input

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: copilot-swe-agent[bot] <198982749+Copilot@users.noreply.github.com>
shunsuke-shimomura added a commit to ut-issl/inka that referenced this pull request Nov 2, 2026
…test

Vite 8.3.5 reads the inline source map with `code.charCodeAt`, but
vitest 5.0.5 hands it a Buffer for modules served from its transform
cache (vitest-dev/vitest#11682). Any error thrown by a reporter then
crashes vitest's error formatting with `code.charCodeAt is not a
function`, which hides the reporter's explicit failures and breaks the
adapter tests that assert on them. The upstream fix (vitest-dev/vitest#11709)
is merged but not released yet, so pin vite through `overrides` in this
package only; the web UI is unaffected.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
@github-actions github-actions Bot locked and limited conversation to collaborators Nov 8, 2026
Sign up for free to subscribe to this conversation on GitHub. Already have an account? Sign in.

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

Reporter errors crash source-map formatting with Vite 8.3.5 (code.charCodeAt)

2 participants