* docs(adr): propose agent AI SDK v7 adoption plan
* fix(ai): bind analyzeResume through aiService in service test
The test destructured analyzeResume as a named export that does not exist; main was red.
* test(agent): keep pure ai helpers real via spread-actual mock factory
* feat(agent): add run guards, patch version guard, run wall-clock timeout
* feat(agent): validate UI messages at the send boundary
* feat(agent): crash-safe draft-row persistence and server-side cancellation
* feat(agent): reap stale run claims at boot, on send, and on thread open
* feat(agent): fresh-document patch output and tiered context pruning
* feat(ai): shared agent tool contracts and message metadata schema
* feat(agent): add per-thread review-patches setting with update endpoint
* feat(agent): gate resume patches behind hmac-signed tool approval
* feat(agent): merge question answers and approval decisions before run claim
* feat(agent): approval ui with composed auto-send and fixture-driven tests
* feat(agent): usage metadata, tool activity cards, smoother streaming
* feat(agent): tool-call repair, input examples, structured step logging
* chore(i18n): translate new agent workspace strings across all locales
* fix(agent): gate stale-run draft cancellation on winning the claim clear
Snapshot streaming drafts before the conditional clear and skip the flip entirely when another reaper or a replacement run already cleared the claim. Also address review nits in eleven locale catalogs.
* fix(agent): flip reaped drafts only when their snapshotted state is unchanged
* fix(agent): address review findings across run lifecycle, context budget, and approval flow
- bind patches to the revision the model read via signed baseUpdatedAt
- claim the run before consuming a continuation; recorded-but-unexecuted approvals retry as pending continuations
- keep run ownership on stop() until cancellation persists; preserve the claim for the reaper when final persistence fails
- estimate tokens without serializing binary attachments (tokenx) and enforce the budget by dropping oldest whole turns
- mark crash-recovered patch results as snapshot boundaries; strip /data prefixes at execution time
- retry failed continuations without regenerate; mount a single AgentChat; disable response controls on read-only threads; freeze review toggle during runs (client+server)
- accumulate usage across continuations and match the SDK's nested usage shape; label-form token strings; reorderable source label; accessible note field; state-neutral web-search label
* chore(i18n): translate revised agent strings across all locales
* fix(agent): harden baseUpdatedAt validation and address review follow-ups
- bundle tokenx in the server runtime dependencies (e2e boot failure)
- strict ISO schema for baseUpdatedAt plus loud executor rejection of unparseable values
- it-IT source label consistency (Fonte)
- prove penultimate-turn retention in the context pruning test
* chore(deps): exempt tokenx from knip for the externalized server bundle
`SectionItemHeader` only rendered its own box when a template opted into
`mainItemHeaderBorder` (only Ditgar did). Everywhere else it walked the
header children and attached the resolved `item-header` style to the first
descendant that happened to be a literal `View` or `InlineItemHeader`.
Sections whose header starts with anything else — certifications, awards,
projects, publications, references — matched nothing, so the style was
silently dropped; stacked headers such as experience matched only their
first row, so a second row went unstyled.
The header now always renders its own `Div`, so `item-header` covers the
whole header row of every section on every template. `Div` rather than
`View` keeps the base row gap the rows used to inherit from the item box,
and Ditgar keeps its tight header via `rowGap: 0` on its own
`sectionItemHeader` slot, so rendered output is unchanged apart from the
newly styled rows. `mainItemHeaderBorder` is now dead and removed.
Fixes#3349
* docs: remove .superpowers
* feat(dsh-plugin): bring the DeepSeek Harness plugin into the monorepo
Moves dsh-plugin-reactive-resume out of its own repository and into
packages/dsh-plugin. It stays a published, public npm package — the only
one here — but now builds, typechecks, tests, and lints under the same
turbo tasks as everything else.
The move pays for itself in the drift guard. Standalone, the plugin kept a
generated snapshot of the tool names scraped from the live server card at
https://rxresu.me, plus a weekly CI job to notice when that snapshot went
stale. Sitting next to packages/mcp, it reads MCP_TOOL_NAME directly, so a
tool rename breaks the prompt guide on the same pull request instead of
days later. The snapshot, the fetch script, and the scheduled job are gone.
packages/mcp gains a ./tool-names export so that import goes through the
public export map rather than another workspace's src.
Also flips autoInstallPeers off. The DeepSeek Harness rc packages declare
peers that are host-supplied and, in one case
(@deepseek-ai/dsh-type-meta), not published at all, so auto-install 404s
the whole workspace. Turning it off drops only optional peers elsewhere;
@neodrag/core was the single hard peer that had been arriving implicitly,
and it is now declared where it is used. Full typecheck and test suites
pass, and pnpm peers check reports nothing new beyond the pre-existing
drizzle-orm range mismatch.
Tests move from test/ to colocated src/*.test.ts and the build output from
lib/ to dist/ to match repository conventions.
* fix(dsh-plugin): ship a bundle manifest and target the current Harness
`dsh plugin add` warned that the package "declares no dsh.bundle — installed
as a plain dependency, not a profile layer", and it was right. Every other
Harness plugin, in-box and third-party, ships a cordis.patch.yml and points
dsh.bundle.patch at it; that declaration is what joins a package to a
profile's bundle stack. Without it the package installed and then sat inert,
and the README's hand-written insert row was a workaround for the gap rather
than the intended way in.
The peer ranges were also a generation behind. They asked for
@deepseek-ai/dsh-mcp-client and dsh-system-prompt at ^0.0.1-rc.1, which
cannot match the 0.1.0-rc.6 a current harness ships, so the plugin could
never have resolved against the thing it targets. Both APIs are unchanged
across the bump — StreamableHttpConfig still takes the same six fields and
PromptSection still takes name/order/text — so this is a range correction,
not a migration.
That bump pays for itself elsewhere. The old generation peer-depended on
@deepseek-ai/dsh-type-meta, which was never published, and working around
that 404 is why merging this package turned autoInstallPeers off for the
whole repository and pulled @neodrag/core in by hand. The new generation
dropped that peer and publishes every other one, so both changes are
reverted and pnpm-workspace.yaml is back to what it was.
Because a bundle patch mounts the plugin the moment it is installed, a
required apiKey would fail config validation and take the profile down
before the user ever had a chance to mint a key. It now defaults to empty
and apply() warns and mounts nothing, matching how dsh-honcho-memory
handles the same problem.
Verified by packing the tarball and installing it into a clean project with
default pnpm settings: it resolves, imports, and reports its exports.
context.issue spreads to { owner, repo, number }, but Octokit v9 requires
issue_number. The request hit /repos/.../issues//labels and returned 404,
so no opened issue was ever labeled.
Section item headers render the title and its trailing date inside a shared
split row styled with `flex-wrap: wrap`. When the title is long the date wraps
onto its own line and left-aligns instead of staying pinned right, which reads
as inconsistent down a list of certifications.
That row had no selector. It is a bare View, so it never reached the semantic
tree: `item-header` matches the box around the row, and the title and date are
text nodes that layout properties do not apply to. There was no stylesheet that
could reach it.
Expose it as `template-part[name="item-header-row"]`, shared by every template
because the row comes from the shared section components. Awards,
certifications, projects and publications are covered; experience, education
and volunteer stack two rows and hand their headers to the
`inline-item-header-*` parts on some templates, so they are left alone.
Readers can now write:
@version 1;
template-part[name="item-header-row"] { flex-wrap: nowrap; }
The sonner wrapper was the only consumer of next-themes; the Base UI toast
that replaced it does not use the hook. knip flagged it as unused, and CI's
`knip --fix` step removed it and then failed `pnpm check` against a lockfile
that still listed it.
Fills the 25 strings added this cycle by the toast migration, the account page
rename, the Custom Styles status labels and the job posting auto-fill, across
all 53 target catalogs. The zu-ZA pseudo-locale is intentionally left empty.
Adds the questionnaire and empty-state primitives and renders the
ask_user_question tool call inline in the chat, so the agent can offer choices
instead of guessing when a request is ambiguous.
Fetching an arbitrary job URL server side meant owning SSRF defence, redirect
and size limits, and per-site scraping quirks. The autofill tool now takes only
pasted text, so the URL input, the fetch path and its MCP annotation are gone.
The sheet gates the call behind a tested AI provider and a minimum paste length
so a stray snippet does not spend an AI call.
The page now holds account-level actions rather than only destructive ones, so
it is reachable at /dashboard/settings/account and presented with a neutral
icon in the sidebar and command palette.
Swaps sonner's toast.success/error/loading/dismiss for the new toast.add({ type,
description }) and toast.close across dialogs, auth pages, the builder, the
dashboard and the applications views. Behaviour is unchanged.
Turbo defaults to ten concurrent tasks and each vitest sizes its pool to the
core count, so a ten-core machine ran roughly a hundred workers and a 1.6s test
blew its 15s budget. Different suites failed on every run. At concurrency four
the whole repo passed five runs straight with no wall-clock cost.
Each case rebuilt the whole spec, which walks every router and resume JSON
schema. The first case already carried a raised 15s timeout and still timed out
on a loaded machine. The spec is deterministic and only read here, so build it
once: the file drops from over 15s to 1.86s.
Without isolation the files in a worker share one module registry, so a
vi.mock of @reactive-resume/env/server in one file leaked into another and
whichever file imported the module first won. Measured on a clean cache,
isolate: false failed four of four whole-repo runs; with isolation, none.
Units that transitively import the validated server env threw at import time
whenever no .env was present, taking out packages/auth and packages/api. Seeding
the three required variables in the shared setup fixes every current and future
caller in one place. Real values still win.
The case passed cursor position 5 into "--resume-", which lands mid-token and
reads as a selector context, so it received the selector list. Every other case
in the file uses source.length.
The expected offset disagreed with its own line and column: line 2 column 17 is
offset 28, which is where `red` starts. Offset 31 pointed at `; }`. The sibling
UTF-16 case in the same file already used the correct arithmetic.
Anchors the offset to the source it must point at so it cannot drift again.
Disabling the textarea for the duration of a response made the browser blur it,
so the caret left the composer on every send and had to be clicked back. send()
already ignores calls mid-stream, so Enter stays a no-op and type-ahead works.
Every step-start part serialises to the same JSON, so the content-derived key
collided for any multi-step assistant message and React warned about duplicate
keys on each incoming chunk. Two identical text parts collided the same way.
Parts are append-only and never reordered by the AI SDK, so the index is stable.
Public resume pages only produced their OpenGraph and Twitter tags client side,
so a shared link had no card at all. The server now injects them into the shell
and swaps in the resume's own title and description.
The lookup is scoped to public, password-free resumes and deliberately avoids
resumeService.getBySlug: that counts a view and would expose a protected
resume's summary to an unauthenticated crawler. User-authored values are escaped
before they reach the HTML, and any lookup failure falls back to the plain shell.
getResumeSocialMeta is shared with the client route head so the two cannot drift.
X reads twitter:* meta tags from the name attribute, not property, so the card
validator reported twitter:title and twitter:description as missing. Also adds
the og:type tag the root head was never emitting.
The static middleware was mounted ahead of the web app fallback, and Hono's
serveStatic resolves "/" to the directory and returns dist/index.html verbatim.
handleWebApp never ran for the root route, so the OpenGraph, Twitter, canonical
and JSON-LD markup it injects was missing in production - fetching
https://rxresu.me/ as Twitterbot returned zero og: tags.
Route "/" explicitly before the static middleware so the injection runs.
Node 24 terminates the whole process on an unhandled promise rejection, so a
single request's stray rejection could take the server down for every user
(as the USER_STOPPED agent-abort bug did). Add a process-level unhandledRejection
handler that logs and keeps serving. Uncaught exceptions are intentionally left
on Node's default crash-and-restart, since process state is unsafe afterward.
Claude-Session: https://claude.ai/code/session_01ULhhLQ24DvnYwzP4afDuye
Stopping (or archiving) an agent run called controller.abort("USER_STOPPED")
with a plain string reason. The AI SDK only recognizes a cancellation when the
reason is an AbortError (err.name === "AbortError" / isAbortError); a bare
string is treated as a real stream error, and its rejection escaped the
background resumable-stream pump and crashed the whole server process with
ERR_UNHANDLED_REJECTION on every user Stop. Abort with a DOMException named
AbortError (label preserved as the message) so the SDK cancels the run
gracefully. Same fix for the USER_ARCHIVED path.
Claude-Session: https://claude.ai/code/session_01ULhhLQ24DvnYwzP4afDuye
- Add an sr-only "Copy secret" label to the icon-only copy button in the 2FA
enable dialog; it was previously announced as an unlabeled button.
- Add a focus-visible ring to template gallery cards. The only ring was gated
on the selected state, so keyboard focus was invisible while tabbing.
- Disable the API-key create submit button while the request is in flight to
prevent duplicate keys from a double-click.
Surfaced by a shadscan UI audit. The remaining ~95 findings were false
positives from the auditor not understanding the pnpm monorepo and the
TanStack Start root-route shell, and were waived.
Claude-Session: https://claude.ai/code/session_01JYTniVDeA56o1kGhdoCUoD
The dependency-update commit bumped @react-pdf/renderer 4.5.1->4.6.0 (textkit
6.3.0->6.4.0) and silently dropped the pnpm patch that overrides font vertical
metrics to prefer OS/2 sTypo* over inflated hhea values, matching browser line
boxes. 6.4.0 did not upstream it, so the semantic-CSS template visual snapshots
(baselined with the patch, maxDiffPixelRatio 0) no longer matched and the E2E
job failed. Re-create the patch for textkit@6.4.0 and re-register it in
patchedDependencies; remove the orphaned 6.3.0 patch.
Claude-Session: https://claude.ai/code/session_01ULhhLQ24DvnYwzP4afDuye
The server PDF preflight spawned a fresh worker per semantic-CSS edit, each
racing a 15s startup deadline to cold-load the ~721kB+5MB PDF runtime. That
load is super-linear in CPU (~3s at 1 vCPU, >15s on a throttled/shared vCPU),
so on a constrained box every edit hit the startup-timeout path and returned
STYLESHEET_PREFLIGHT_WORKER_FAILED. Since the service only advances the applied
stylesheet when preflight passes, custom styles never applied and the editor
stuck on Checking.
Warm one worker at boot and reuse it (message-based input, respawn on
crash/timeout), so the cold load is paid once instead of per edit. Raise the
render deadline 5s->30s (a rich resume renders ~5-18s on a slow box) and the
readiness ceiling to 120s so the one-time warm completes even when throttled.
Surface worker load failures instead of an unhandled-rejection crash, and log
runner-side failure paths so the previously opaque failure is diagnosable.
Verified in node:24-slim under --cpus=0.25/0.35/0.5: all reused requests pass.
Claude-Session: https://claude.ai/code/session_01ULhhLQ24DvnYwzP4afDuye
Concurrent compile requests from editor intelligence were rejecting
in-flight edit compiles as stale, and the store swallowed that rejection
without leaving compiling. Resolve all compile results and surface
compile failures as an error status so styles can apply again.
Co-authored-by: Cursor Agent <cursoragent@cursor.com>
Add comprehensive Application Tracker REST and MCP coverage, document the MCP workflow, add Markdown/ActionLint checks, bump the release version, and fill all extracted translations.
Amruth Pillai
