Keep stable v0.7 guidance available while the client-entry layout is previewed for v0.8. Existing reference and migration URLs redirect to the release that owns their content.
* feat(workspaces): gate fork repository automation
Cross-repository change requests can execute contributor-controlled package hooks even when the setup command comes from the base repository. Persist the checkout provenance and require one explicit workspace-level setup action before repository automation runs.
* fix(workspaces): clarify blocked script errors
* fix(app): keep long toasts readable
* test(server): expect fork repository provenance
* test(server): make setup marker portable
* test(app): retry transient sidebar layout
* feat(plugins): phase 1 split runtime entries after green gate
The compiler now builds explicit client and server entries, so runtime boundaries are source-owned instead of registration-name filtering. Compiler and runtime tests prove the entry split, directory and suffix boundaries, client Node import errors, migration failure, and client-only operation without a subprocess.
* feat(plugins): phase 2 run client entries after green gate
Explicit client and server contexts replace the mixed plugin context. The app runner owns every client registration and its idempotent removal, including late contributions and composer pills; focused contribution tests prove the removal contract, and all five migrated examples reached running on the isolated worktree daemon.
* feat(cli): phase 3 scaffold runtime split after green gate
The generated project demonstrates the required shared RPC contract, server handler, client surface, and sidebar wiring. The scaffold test and an isolated init, typecheck, install, and running check prove the phase acceptance path.
* docs(plugins): phase 4 publish migration after green gate
Gate: current docs and the paseo-plugin skill describe only explicit client/server entries. The standalone migration guide maps every former registration, and migration-doc.test.ts proves every client add* method remains represented.
* fixup! feat(plugins): phase 1 split runtime entries after green gate
Gate: compiler 13, runtime 24, typecheck, lint, and format passed. The dependency fixture proved node_modules/client was incorrectly treated as a plugin boundary.
* fixup! feat(plugins): phase 2 run client entries after green gate
Gate: client runtime and registry tests, typecheck, lint, format
* fixup! docs(plugins): phase 4 publish migration after green gate
Gate: migration docs test, paragraph audit, typecheck, lint, format
* fixup! feat(plugins): phase 1 split runtime entries after green gate
Gate: compiler organization tests, typecheck, lint, targeted format
* fixup! feat(plugins): phase 2 run client entries after green gate
Gate: shared SDK tests, runtime tests, typecheck, lint, targeted format
* fixup! feat(cli): phase 3 scaffold runtime split after green gate
* fixup! docs(plugins): phase 4 publish migration after green gate
* fixup! feat(plugins): phase 1 split runtime entries after green gate
* docs(plugins): rewrite public plugin docs for runtime entries and mobile guardrails
* fixup! feat(cli): phase 3 scaffold runtime split after green gate
The scaffold now keeps DOM globals out of the program and declares only window.open inside client/web.ts. The scaffold test invokes tsc in a fresh process, proves the generated project passes, and proves a stray document access fails.
* fixup! docs(plugins): phase 4 publish migration after green gate
The internal guide and plugin skill now forbid both the DOM lib and triple-slash DOM references. They direct web adapters to declare only the globals used by client/web.ts.
* fixup! feat(plugins): phase 1 split runtime entries after green gate
Removing the obsolete Babel parser changes the locked npm dependency graph. The macOS Nix desktop gate reported the new fixed-output hash, which this commit records.
* fixup! feat(plugins): phase 1 split runtime entries after green gate
Reject relative imports that escape the plugin root while continuing to skip resolved node_modules internals. The compiler regression test proves the escaped import is rejected and the dependency-internal fixture remains accepted.
* fixup! feat(plugins): phase 1 split runtime entries after green gate
Classify absolute imports and reject plugin-authored paths that escape into node_modules. Dependency internals remain exempt based on their importer path. Red-first compiler regressions cover both bypasses, and the existing dependency fixture remains green.
* fixup! feat(plugins): phase 1 split runtime entries after green gate
Classify canonical esbuild resolution results so dependency-relative and symlink imports cannot escape client/server boundaries. The two compiler regressions failed before the fix and pass afterward.
* test(ci): synchronize flaky state transitions
Wait for repository watcher registration before emitting buffered ref events, let sidebar order polling retry unlaid-out rows, and wait for the inactive browser parking state before screenshot capture. These changes directly address the three observed CI failures; the affected server, Playwright, and desktop browser tests pass locally.
* test(ci): synchronize Mermaid completion layout assertion
The Playwright streaming acceptance test exposed a completion remount between visibility and layout sampling. Check the existing completion promise around the measurement so that transition is not reported as diagram loss.
* fixup! feat(plugins): phase 1 split runtime entries after green gate
Resolve bare package imports before boundary classification so symlinked dependency entries cannot expose server modules to the client bundle. The focused regression proves the bypass and the compiler file passes 20/20.
* fixup! feat(plugins): phase 1 split runtime entries after green gate
Allow canonical paths only within a matching linked package root while preserving runtime and containment checks for imports that leave it. The linked-dependency regression fails before the fix and compiler tests pass 21/21 afterward.
* fixup! feat(plugins): phase 1 split runtime entries after green gate
Reject matching package manifests that contain the plugin or live inside it, so linked-dependency roots cannot exempt plugin or workspace files. The ancestor-manifest regression fails before the fix and compiler tests pass 22/22 afterward.
* fixup! feat(plugins): phase 1 split runtime entries after green gate
Validate plugin-local lexical boundaries before canonical linked-root exemptions and require remembered-root imports to originate inside that root. The linked-root bypass regression fails before the fix and compiler tests pass 23/23 afterward.
* feat(app): make import session a first-class entry point
* feat(server): improve import session discovery resilience
Search up to 500 persisted candidates per provider and isolate listing failures behind eight-second deadlines so healthy results remain usable. Preserve native persisted title precedence while keeping discovery read-only.
Parsing now continues through each transcript so native titles and the true last prompt are retained; the old early exit made lastPromptPreview equal firstPromptPreview. A cheap record prefilter avoids JSON.parse for unrelated lines during the full scan.
* feat(app): searchable, grouped import session sheet with per-provider errors
The sheet now carries a debounced search field gated on the host's
importSessionSearch feature, grows its per-provider page on demand
(15 -> 45 -> 90 -> 200) instead of capping at the 15 most recent, and groups
unscoped rows by folder under the project name the app knows for that
directory. Worktrees of one project would otherwise all head up as the same
bare name, so a directory below its project root carries the path under that
root beside the name.
A workspace-scoped sheet says so and offers "Show all sessions". Importing a
row from another directory once that is on drops the workspace from the
request, which the daemon would otherwise reject, and routes the user to the
agent's own workspace rather than opening a dead tab in the current one. That
decision lives in resolveImportTarget: a scoped listing is trusted wholesale
because the daemon already matched it realpath-aware, which the client cannot.
A provider that fails now gets its own error row with a Retry that refetches
only that provider, sourced from both a failed request and the daemon's
providerErrors. React Query no longer retries these, so the header spinner
settles once every provider has answered or given up (#2512).
* test(e2e): import session flow with fixture Claude sessions
* fix(app): shorten import sheet scope copy so it fits at phone width
At 390px the scoped subtitle rendered "Sessions in this wor..." — the label,
its action, and the sheet's refresh and close buttons do not share one line at
phone width. The label already sits under the sheet title "Import session", so
"Sessions in this" was repeating context the header had: it becomes "This
workspace" and the action becomes "Show all". The unscoped label still names
the host, which nothing else on the sheet does.
* refactor(app): flat recency-ordered import list with per-row folder label
Grouping by folder buried recency: a busy directory's block of fifteen pushed
a two-minute-old session in a quiet directory below the fold, and the axis a
user is scanning on is when they last touched the session, not where it lives.
The unscoped sheet is now one list in the order aggregateSessionEntries already
produces, and each row carries its folder as a muted third line.
resolveDirectoryLabel stays — it is what keeps two worktrees of one project
apart — and formatDirectoryLabel renders its two parts as the single ellipsised
line a row has room for. A scoped sheet lists one directory, so its rows carry
no folder line at all.
* fix(app): close import sheet before refreshing sessions
Provider-list invalidation waited for every active query before the sheet could close. Mark the disabled picker queries stale after close and navigation so reopening refreshes without delaying the imported agent.
* fix(server): reuse workspaces for session imports
Host-wide imports should join the oldest active workspace for their directory, or restore its archived workspace, instead of minting duplicates. Rollback remains limited to workspace records created by the failed import.
* test(e2e): express the import session journey through a helper DSL
* test(e2e): align sidebar journeys with import entry point
A repository-controlled core.fsmonitor command can execute when the daemon refreshes Git state. Opening or reconciling a malicious checkout therefore allows code execution as the daemon user. Force core.fsmonitor=false on daemon-owned Git invocations before repository config is considered.
Plugins can transform live timeline source items at render time, append durable plugin rows from daemon handlers, and register client-only slash commands in the composer.
Related to #3940 and #3952.
* feat(claude): add Fable 5.1 model
Expose the new 1M-context model through the curated Claude catalog and normalize its dated and context-suffixed runtime IDs back to one canonical entry.
* fix(claude): group Fable models in catalog
* chore(release): cut 0.7.1
* fix(nix): refresh 0.7.1 dependency hash
* test(cli): follow package version in plugin scaffold
* test(cli): keep Claude catalog checks extensible
* fix(ci): eliminate recent cross-platform flakes
Synchronize tests on owned state instead of elapsed time or transient identities. Preserve file rename collision safety when platform file IDs are unavailable.
* fix(ci): close follow-up synchronization gaps
Keep fetches in flight through buffered-event classification, scope mock agents to navigation tests, and wait for live workspace hydration before opening agent routes.
* fix(ci): synchronize browser lifecycle boundaries
* feat(plugins): support prepared Git installations
Let Git plugins declare explicit preparation commands without making package installation the default. Keep installation transactional, make plugin trust explicit, and use one canonical source reference for monorepos.
* fix(plugins): parse Windows source references
* fix(server): keep daemon alive when JSONL RPC stdin hits EPIPE
A dead Claude/OMP child currently turns stdin.write() EPIPE into
uncaughtException and kills the daemon. Catch the write, listen for
stdin errors, and close that transport instead.
Closes#4047
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
* test(server): reproduce JSONL RPC stdin EPIPE
* test(server): gate native EPIPE fixture on POSIX
---------
Co-authored-by: Claude Sonnet 5 <noreply@anthropic.com>
Co-authored-by: Mohamed Boudra <boudra.moha@gmail.com>
Track each foreground run start as pending, started, or failed so a previous turn error cannot reject a newly accepted prompt. Includes a regression for resuming after a failed turn.
* feat(hub): make daemon authority explicit
Connecting a daemon establishes presence without granting execution. Hub authority now uses semantic session permissions, with workflow execution granted explicitly and inspectable through the normal session protocol.
Attention frames remain part of the existing agent and terminal subscriptions, keeping session output solicited.
* test(server): match sentence-case tool labels
Mid-stream silence clearing can leave a non-final partial whose segment will never commit. Treat those partials as UI-only once all requested commits settle so retries of the same recording can complete instead of timing out forever.
Keep trusted-client and Hub behavior stable while replacing RPC-name grants with a semantic authorization boundary. This establishes the operation inventory needed for future principal, pairing, and resource-scoped access work.
xAI moved Grok accounts to unified billing. The plain /v1/billing
response now returns a zeroed legacy monthly shape (monthlyLimit.val 0),
which rendered as "0 credits left" with no bar in Settings -> Usage.
Fetch /v1/billing?format=credits — the same query the Grok CLI /usage
command uses — and map creditUsagePercent + currentPeriod onto a usage
window, matching the Claude/Codex weekly window convention. The legacy
monthly-credits balance path is unchanged; unified-billing responses
omit monthlyLimit/used, so it naturally yields no balance.
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
* fix(opencode): reuse servers across agent sessions
Paseo's always-present per-agent environment forced every OpenCode session onto a dedicated helper, multiplying cold starts and memory use. Route ordinary sessions through one shared helper while a daemon-owned session bridge preserves exact environment and caller-scoped tools; keep custom process environment and MCP configuration isolated.
* fix(build): bundle the agent bridge from existing dependencies
The helper package duplicated Zod and changed the fixed-output desktop dependency graph even though its tool wrapper was identity-only. Use the server's existing Zod runtime directly so packaged daemons can materialize the bridge without extra traced dependencies.
* refactor(server): isolate provider runtime composition
Keep provider-specific lifecycle and tool-catalog wiring behind an agent-owned runtime boundary so the generic daemon bootstrap remains provider-agnostic. The runtime now owns bridge startup, snapshot injection, and idempotent cleanup.
* fix(desktop): prebundle the agent bridge plugin
Packaged daemons run from an application archive where esbuild cannot spawn its binary. Produce the self-contained plugin during the server build and make compiled runtimes read that artifact; source-mode tests retain a gated compiler fallback.
* fix(app): preserve archived agent navigation
Make explicit archived-agent opens one atomic layout command and defer them until persisted layout hydration. Use the provider's numeric active-session sentinel so recovery can complete.
* fix(app): preserve timeline resume during hydration
Paint persisted canonical rows with the current live head and retain their resume cursor. This prevents tail replay without dropping stream updates that arrive while replica storage loads.
* fix(app): serialize replica reads with cache writes
The initial retained poll could run before repository identity discovery settled, routing every workspace through the legacy per-target path. Defer each unresolved target independently and rejoin the shared polling grid once its identity is ready.
MCP and extension initialization can keep provider control-plane requests pending beyond 30 seconds even when the response will arrive. Use a configurable 60-second provider default and include the active RPC phase in timeout errors.
* fix(dictation): preserve audio queued at stop
Stopping capture disabled native recording before the final microphone buffer had crossed the bridge, so accepting immediately could omit the end of a dictation. Wait for the native producer to publish its final buffer before completing stop, and serialize stop with restart and destruction.
* fix(voice): serialize capture lifecycle
A capture start already awaiting permission could resume after teardown, and repeated destroy calls did not share completion. Order capture lifecycle operations behind one promise and build the native module declarations before clean-CI typechecking.
* fix(dictation): drain queued web capture before finish
* fix(dictation): await in-flight segment commits
* Revert "fix(dictation): drain queued web capture before finish"
This reverts commit 5168a6554d.
* Revert "fix(voice): serialize capture lifecycle"
This reverts commit 00fd3e3556.
* Revert "fix(dictation): preserve audio queued at stop"
This reverts commit ec087dc9cd.
Paseo forced signing off for commits created through the app and squash merges. Let Git apply the user's configured policy; the existing command timeout still bounds interactive signer waits.
Co-authored-by: paseo-bot[bot] <266920839+paseo-bot[bot]@users.noreply.github.com>
* feat(plugins): add React Native UI primitives
Give client plugins host-owned adaptive UI without exposing app internals. Preserve plugin runtime contexts when modal content moves through the compact portal host.
* docs(plugins): clarify host UI reference
* refactor(plugins): harden host UI boundary
Replace mocked pass-through coverage with a real browser plugin journey, keep host runtime ownership inside the evaluator, preserve toast context across compact portals, and expose modal content as a compound component.
* fix(plugins): defer native UI presentation runtime
Loading the plugin evaluator initialized the modal and toast presentation stacks even when no plugin requested host UI. Keep the capability stable while delaying platform-heavy presentation code until Modal.Content renders.
* refactor(plugins): move modal metadata to root
Keep Modal.Content body-only while the top-level Modal owns the host-rendered title and leading icon. This leaves presentation metadata stable as the compound API gains future regions.
* Batch GitHub PR status polls into aliased GraphQL requests
The PR status poller ran one `gh pr view` (GraphQL-backed) plus a
per-PR facts query for every workspace on its own timer, so cost
scaled linearly with workspace count and exhausted the 5,000 pts/hour
GraphQL quota machine-wide past ~50 workspaces (#3587, #2470).
Replace the per-target timers with a global scheduler that snaps poll
targets to a shared 5s grid and flushes all due targets together:
- One aliased discovery query per batch of up to 25 targets (grouped
by GH host) resolves PR identity, state, mergeability, and the merge
facts that previously required a second GraphQL call per PR.
- A second aliased query fetches statusCheckRollup contexts only for
the PRs that matched. Checks for merged/closed PRs are frozen and
cached by repo#number@headOid, so idle workspaces settle to the
discovery query alone.
- A failed batch backs off every target in it with one request spent;
healthy aliases are recovered from the body gh still prints on
stdout when a single alias errors (ForgeCommandError now carries
stdout for this).
- Targets whose origin slug/host is still resolving, and fork repos
whose PR lives in the parent, fall back to the legacy per-target
path for that cycle. Unowned lookups filter out other forks' PRs
that share a branch name.
Fixes#3587. Addresses the API budget in #2470.
* fix(server): reserve the shared GitHub polling budget
---------
Co-authored-by: Mohamed Boudra <boudra.moha@gmail.com>
* fix(server): keep Git pressure from stalling daemon
Git process starts were refilled synchronously, so slow process spawning could monopolize the main loop long enough to trip connection liveness checks. Admit one process per event-loop turn and expose refresh/fetch provenance in runtime queue metrics.
* perf(server): avoid delaying isolated Git commands
Start the first command immediately when the scheduler is idle. Backlog admission and completion-driven refills still yield between spawns, preserving daemon liveness without adding a turn to normal sequential Git traffic.
* fix(server): keep Git admission outside fake clocks
Subprocess admission is event-loop ordering rather than simulated wall-clock behavior. Use the native Node immediate so service tests can fake debounce timers without deadlocking queued real Git processes.
* refactor(server): bind Git command provenance explicitly
* test(server): expect bound Git fetch runner
* Add steering support for ACP providers that indicate support
* fix(acp): keep steering inside the admitted turn
---------
Co-authored-by: paseo-bot[bot] <266920839+paseo-bot[bot]@users.noreply.github.com>
Plugin client bundles bypass Metro, so authored async callbacks reached Hermes without Metro's syntax lowering. Plugin connection inputs also selected bottom-sheet behavior from device form factor instead of their actual render scope.
* feat(plugins): add timeline item contributions
Keep canonical daemon history unchanged while plugins project and render custom timeline items in clients. Live matches refresh authoritative projected history before replacement so lifecycle deltas stay collapsed.
* fix(nix): update npm dependency hash
The plugin SDK lockfile change invalidated the fixed-output npm dependency derivation. Use the hash calculated by the macOS Nix build for the current lockfile.
Make every trusted Git repository directly distributable without introducing a registry or package manager. Managed checkouts retain their resolved commit, track branches explicitly, and stage validation before activation so failed updates restore the running version.
* chore(server): upgrade Claude Agent SDK to 0.3.246
Bumps @anthropic-ai/claude-agent-sdk from 0.3.220 to 0.3.246, keeping the
pinned exact version so fresh installs stay reproducible.
The SDK's peer requirements are unchanged (@anthropic-ai/sdk >=0.93.0,
@modelcontextprotocol/sdk ^1.29.0, zod ^4.0.0) and already satisfied, so
nothing else in the tree moves.
Nothing Paseo imports was removed or narrowed. The type surface Paseo
touches is additive; the removed ExitReason member
'bypass_permissions_disabled' and the OnElicitation/OnUserDialog signature
changes are unused here.
* chore(nix): refresh npm deps hash for the SDK bump
---------
Co-authored-by: paseo-bot[bot] <266920839+paseo-bot[bot]@users.noreply.github.com>
Provider callbacks can be canceled without a user response. Publish a terminal resolution for that path so every surfaced request has a matching lifecycle event, while preserving the existing interruption cleanup.
* feat(protocol): add forge check presentation metadata
Expose optional rawStatus, isManual, and requiresAction fields while keeping older check payloads backward-compatible.
* fix(gitea): model Actions jobs as tasks
Gitea's /actions/tasks endpoint returns individual jobs rather than workflow runs. Preserve sibling tasks from the latest execution, avoid collapsing same-name jobs, and enrich blocked tasks with run-level approval metadata.
* feat(app): distinguish forge check presentation states
* fix(gitlab): preserve pipeline job semantics
* refactor(app): centralize forge check presentation policy
Use one canonical category order, count shape, visual policy, and translated copy mapping across sidebar, hover-card, PR-pane, and GitLab pipeline surfaces.
* refactor: consolidate forge check handling
* feat(forge): auto-attach pasted forge references
* docs(compat): correct forge rollout metadata
* refactor(forge): model check nuances as traits
* fix(forge): harden checks from branch review
Structural Gitea Actions shadow dedup by run id (verified upstream),
rerun grouping without workflow_id, shared trait/status constants and
Gitea state mapping in the protocol package, all cloud host aliases in
pasted-link matching, logged glab job-detail failures, concurrent
approvals/pipeline fetch, checks aggregate from the same pipeline as
the job list, and removal of the unread check-level rawStatus. The
released pipeline-level rawStatus is staged for removal under
COMPAT(pipelineRawStatus).
* fix(gitea): detect fork-PR Actions approval hold as action_required
gitea.com represents a fork PR awaiting Actions approval as a run in
status "waiting" with no tasks and pending shadow commit statuses; it
does not expose need_approval. The previous mapping only set
action_required from a task with blocked+need_approval, so the gate
never surfaced (0 tasks -> early return; need_approval absent). Detect
the hold from /actions/runs: a waiting/blocked run whose PR is
cross-repo (head.repo.id != base.repo.id), or one that does expose
need_approval (self-hosted). Tag the run's pending shadow statuses
action_required. Runs are only fetched when a pending shadow has no
covering task, keeping the common path at two requests. Verified live
against gitea.com fork PR (run 715170, waiting, cross-repo).
* fix(gitea): resolve cross-fork PRs by head sha
The current-PR matcher gated candidates on the checkout owner
(giteaHeadOwner === expectedHeadOwner), which excludes every fork PR:
its head carries the fork owner while the checkout is the base repo, so
a fork PR - the only shape that can be approval-held on Gitea - never
resolved and its checks never showed in-app. GitLab has no such gate
(it resolves fork MRs from the target project by source branch). Accept
a recent PR whose head sha equals the workspace head sha regardless of
owner; the sha is an unambiguous identity, so the anti-collision owner
gate still applies when no sha match is available. Verified live against
gitea.com fork PR #6 (resolves + renders action_required).
* fix: resolve forge rebase drift
* fix: type forge auto-attach boundary
* fix(i18n): add Korean check status copy
---------
Co-authored-by: Mohamed Boudra <boudra.moha@gmail.com>
* Stream assistant text at a steady rate instead of in arrival lumps
Assistant text arrived in lumps and was painted that way. The daemon
coalesces one message per 60ms window, and that window carries however
many characters the model produced in it, which swings by an order of
magnitude within a turn. Painting each delta as it landed made the size
of those lumps visible.
Three changes:
- The agent stream coalescer is now leading + trailing. The first delta
after an idle window flushes synchronously instead of waiting a full
window, matching the invariant TerminalOutputCoalescer already
documents. The window stays at 60ms; shrinking it would raise message
rate on a loop that already contends with terminal frames and relay
encryption, without making the stream smoother.
- The client reducer queue commits on a frame instead of an arbitrary
48ms timer, with the timer kept as a ceiling so a hidden tab still
advances.
- Assistant and reasoning text are revealed at a rate derived from the
backlog rather than from arrival timing. Arrival changes the target,
not the rate. The store keeps the full text and only the rendered
slice is paced, so copy, selection, the chat outline, and scroll
geometry all stay consistent with what is on screen.
The leading flush means a burst's first chunk lands as its own timeline
row. The projection already merges contiguous same-turn rows and clients
read the projected timeline, so history is unchanged; the server tests
that assert on raw rows were updated for the extra row.
Measured on Expo web against a local dev daemon. Mock bursty-stream
provider, 6s samples, paint-on-arrival vs paced: frames that advanced
the text 55 -> 244, chars-per-frame CV 2.66 -> 1.44, gap p50 100ms ->
17ms. Real Claude Haiku agent over 52s: 84% of frames advance the text,
CV 1.29, gap p50 and p95 both 17ms (one frame at 60Hz).
Adds a bursty-stream mock model with seeded burst sizes, a gated browser
perf spec, and docs/agent-stream-performance.md.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
* Address review: safe reveal boundaries and a testable seam
Two findings from the PR review.
Unicode boundaries. Slicing the revealed text at an arbitrary index could
strand a lone surrogate, which paints as a replacement glyph, or split an
emoji ZWJ sequence so a family briefly rendered as its members. Both were
visible for a frame or two mid-stream — the exact flicker the paced reveal
exists to remove. The rendered cut is now pulled back off surrogate pairs,
combining marks, variation selectors, skin tone modifiers, and ZWJ joins.
The reveal counter stays monotonic and only the painted boundary moves, so
a long cluster cannot stall the reveal.
Test placement. The hook test used JSDOM, @testing-library renderHook, and
monkey-patched requestAnimationFrame, all of which docs/testing.md rules
out; it prescribes fixing the missing seam instead. The reveal policy is
now a pure state machine in text-reveal.ts (begin/retarget/advance/
complete/visible) covered by plain unit tests with no environment, and the
hook is reduced to the frame-clock binding. Rendered behavior is left to
the browser spec, which no longer carries frame sampling and statistics in
its body — those moved to e2e/support/helpers/stream-smoothness.ts.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
* Measure total painted text, not the tail message
The smoothness sampler read the length of the last assistant-message
element. A real turn emits many assistant messages, so that element keeps
changing identity and its length is not monotonic — handovers read as
resets, and a stream that was plainly moving on screen reported almost no
growth at all. Sampling the total across every assistant message fixes it.
This invalidated the numbers previously recorded for the real-agent run,
so the doc table is replaced with a fresh A/B taken with the corrected
sampler against a Claude Haiku agent: frames that advanced the text 6% ->
87%, chars-per-frame CV 4.11 -> 1.86, gap p50 317ms -> 17ms, p95 383ms ->
17ms. The paced column also reported zero Unicode defects across a stream
containing emoji, which exercises the reveal boundary clamp end to end.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
* perf(agent-stream): harden paced text reveal
Use standards-compliant grapheme segmentation when available and preserve full arrivals on older runtimes. Cap reveal commits at 60Hz and measure the real browser rendering path so high-refresh displays stay smooth without rendering at hardware frame rate.
* Hold back trailing fragments split by arrival boundaries
The reveal boundary is not the only place a grapheme gets cut. The
daemon's coalescing window ends a delta wherever the window closes, so a
caught-up reveal can be handed half a cluster and paint it. Measured on a
mixed-script Haiku stream: five frames in ~350 ending in a single regional
indicator, i.e. half a flag on screen for a frame.
While the turn is streaming, a trailing fragment that cannot stand alone —
a dangling high surrogate, a trailing joiner, an odd regional indicator —
now waits for the rest of its cluster. Once the turn ends everything
received is painted, because nothing more is coming. Complements the
grapheme-boundary clamp, which handles the reveal cut; this handles the
arrival cut.
The same stream now measures one such frame in ~400 rather than five. The
remainder is a flush boundary that coincides with the end of an assistant
message, putting the two halves in different messages, which the client
cannot rejoin — that one belongs in the coalescer.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
Co-authored-by: Mohamed Boudra <boudra.moha@gmail.com>
Mohamed Boudra