807cce6a hides the "Show setup" action when there is no workspace setup
snapshot. That broke the case where setup genuinely ran on a different
WebSocket session than the one the browser reconnected with: the snapshot
was cached per-Session, so discovery returned null and the action stayed
hidden.
Share the setup snapshot cache at the websocket-server level and inject
it into every Session. On focus, the workspace screen now requests
fetchWorkspaceSetupStatus when it has no local snapshot; workspaces with
no setup history still return null, preserving 807cce6a's intent.
mcp-server tests invoked registered tools via .handler, but the MCP
SDK exposes the callable as .callback. model-resolver now passes
force: false to fetchModels; widen the spy matcher to match.
Scripts launched from the UI now run in a dedicated terminal per script
(titled after the script) and reuse the same terminal across re-runs.
Script lifecycle is driven by a new onCommandFinished signal sourced from
OSC 633 command-finished events emitted by the zsh shell integration, so
foreground script exit updates status without killing the terminal.
Script payloads now carry terminalId so the workspace can surface the
associated tab from the scripts dropdown.
Aggregates inbound message counts, bytes, handler timings, and
agent_stream breakdowns over a rolling window and flushes via logger.
Gated on runtimeMetricsIntervalMs config; disabled by default.
CheckoutDiffManager.refreshTarget was calling workspaceGitService.getCheckoutDiff
without forcing a cache bypass, so watch-fired refreshes within the 15s consumer
TTL returned the pre-mutation cached diff — fingerprint matched, no
checkout_diff_update was emitted, and the sidebar diff appeared frozen while an
agent edited files. Route watch-fired refreshes through the service with
{ force: true, reason: "working-tree-watch" } so they always recompute.
Pi's streamHistory only emitted text, so resuming a session lost tool
calls, tool results, and bash executions from the replayed timeline.
Iterate session.messages to also emit running/completed/failed tool_call
items and bashExecution shell calls, and tag user messages with a stable
messageId so hydrateTimelineFromLegacyProviderHistory can dedupe.
Cover it with a real OpenRouter-backed e2e test that runs a bash tool
call, resumes via PiDirectAgentClient.resumeSession, and asserts the
replayed timeline contains the user message, assistant message, and
completed shell tool_call with the expected output.
Centralizes refresh orchestration in WorkspaceGitService and GitHubService:
server coalesces, deduplicates, and throttles git/GitHub reads; clients peek
snapshots (cold reads fall through to a one-shot fetch) and never drive
refreshes. Introduces a two-TTL model (consumer vs internal), force-emit
bypass so fingerprint-matched refreshes still propagate, notifyGitMutation
wiring at ~13 action handlers, adaptive GitHub polling based on check state,
and checkout_status_update / prStatus push-to-cache wire messages. Fixes the
stale diff-stat bug and the inverted base-branch diff direction, and flips
next-action affordances promptly after commit / PR merge.
Replaces the per-client stale/visibility policy with one shared
presence cascade so a backgrounded Firefox tab can no longer steal
in-app notifications from an active Electron desktop window.
Server (`agent-attention-policy.ts`): exports `PRESENCE_THRESHOLD_MS`
(3 min) and `computeNotificationPlan({ allStates, agentId, reason,
nowMs })` returning `{ inAppRecipientIndex, shouldPush }`. Cascade:
focus suppression requires presence; otherwise the most-recent
present client wins (lower-index tiebreak); push fires only when no
client is present, and never on `error`. `broadcastAgentAttention`
calls the policy once per event.
Electron (`daemon-manager.ts`, `idle.ts`): registers
`desktop_get_system_idle_time` returning `powerMonitor
.getSystemIdleTime() * 1000`. Renderer wrapper
`getDesktopSystemIdleTimeMs()` is best-effort: rejected IPC, null,
NaN, Infinity, or negative values return null without throwing.
Hook (`use-client-activity.ts`): Electron-only effect polls every 5s
and updates `lastActivityAtRef` only forward, so OS-wide activity
keeps a backgrounded Electron window present while in-window events
still update immediately.
`ClientHeartbeatMessageSchema` is unchanged; old/new clients and
daemons remain wire-compatible.
Tests: 52 focused tests across policy unit, dispatch, e2e, hook,
desktop wrapper, and permissions. Phase-3 dispatch test +
Phase-5 stale-poll test verified by mutation (breaking the focus
check / the `>` guard makes the exact regression fail).
normalizeCheckoutPrStatusPayload always emits the full wire shape now
(number, repoOwner, repoName, isDraft, checks, checksStatus,
reviewDecision) — the workspace-git-watch test's toEqual assertion
still had the pre-PR-pane shape and was failing.
Update the expected object to match the actual wire contract.
Branch switcher listed remote-only branches as if they were local, then
threw "Branch not found" when clicked because the checkout path only
verified `refs/heads/<name>`. Explicit `origin/<name>` input went the
other way — verify passed, but checkout produced a detached HEAD.
Split the concerns into three layers:
- Enumerate: `listBranchSuggestions` returns provenance (`hasLocal`,
`hasRemote`) so the wire payload no longer loses origin-only info.
Fields are optional on `branchDetails` for backward compat.
- Resolve: new `resolveBranchCheckout` normalizes `origin/<name>` →
`<name>`, verifies explicit `refs/heads/` and `refs/remotes/origin/`
paths (no tag-DWIM), returns a tagged
`{ local | remote-only | not-found }` result. Policy (local wins,
origin fallback) lives here.
- Act: `checkoutExistingBranch` no longer does implicit DWIM. Local →
`git checkout <name>`. Remote-only → `git checkout -b <name> --track
origin/<name>`. Not-found → throws the existing
`Branch not found: <name>` string. Response carries optional
`source: "local" | "remote"`.
`validateBranch` now routes through the same resolver so both paths
agree on what a branch is.
Also fixes a related stale-base diff bug: worktrees created off
`origin/main` could show phantom "added" commits when the local `main`
lagged. `resolveBestComparisonBaseRef` is now origin-first for
diff/status. Merge still uses the "most-ahead" helper so unpublished
local commits aren't lost. `doesGitRefExist` tightened to only swallow
exit 1 (missing ref), propagating other git failures.
Introduce a PR tab in the explorer sidebar that shows status checks,
reviews, and the merged timeline for the current branch's pull
request. Extend checkout PR status with number, draft flag, repo
identity, and per-check workflow/duration; add a new
pull_request_timeline request/response pair backed by a GraphQL
query; enrich the workspace git snapshot to carry the fields through
to the client; and extend the checkout-git-actions store so timeline
queries are invalidated alongside status when mutations land.
Pass the GitHub service into pullCurrentBranch, pushCurrentBranch,
createPullRequest, branch setup, and worktree archival, and invalidate
the cached PR status / repo view whenever a mutation succeeds. Have
mergeToBase return the worktree cwd it actually touched so the session
can invalidate the base checkout when the merge lands there.
Plumb updatedAt through GitHub issue/PR search and committerDate through
branch suggestions so the new-workspace picker interleaves branches and
PRs sorted by last activity. Debounce the search query so remote GitHub
searches are driven by what the user typed rather than a static list.
Pi's ACP binary streams 200+ redundant tool_call_update events per tool call,
bypassing the server coalescer and flooding the app with synchronous state
updates. Two-layer fix:
Server: extend AgentStreamCoalescer to handle tool_call items via last-write-wins
per callId (200ms window, up from 33ms). Terminal statuses flush immediately.
App: mergeToolCallDetail and appendAgentToolCall now return existing references
when fields are unchanged, preserving React.memo equality checks.
* feat(server): replace Pi ACP integration with direct SDK provider
Replace the ACP-based Pi provider (which spawned a `pi-acp` subprocess
and talked JSON-RPC over stdio) with a direct in-process integration
using `@mariozechner/pi-coding-agent` as a library.
The new `PiDirectAgentClient` and `PiDirectAgentSession` use the Pi SDK
directly — creating sessions via `createAgentSession()`, managing models
via `ModelRegistry`, and mapping Pi's `AgentSessionEvent` stream to
Paseo's `AgentStreamEvent` types including:
- Thread/turn lifecycle events
- Assistant text and thinking/reasoning deltas
- Tool execution (bash, read, write, edit, find, grep, ls) with arg
caching across start/update/end events
- Compaction events
- Session persistence and resume via `SessionManager`
Thinking levels flow through the existing `thinkingOptionId` /
`setThinkingOption()` infrastructure, not through features.
Includes 9 real e2e tests against a live Pi agent covering tool calls
(bash/read/write/edit), reasoning chunks, session persistence/resume,
model listing, and thinking option management.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix: update lockfile signatures and Nix hash
* refactor(server): improve Pi direct provider code quality
Address coding standards violations in pi-direct-agent.ts:
- Replace all `as Record<string, unknown>` / `as any` / `as unknown as`
casts with properly typed boundaries using Pi SDK's exported tool
input types (BashToolInput, ReadToolInput, EditToolInput, etc.)
- Introduce named interfaces at all boundaries: PiPromptPayload,
ToolCallOutputSummary, PiModelReference, PiPersistenceMetadata
- Parse tool args once at boundary into typed discriminated union,
then trust types internally — no more typeof probing in mappers
- Break all dense ternary chains into named steps / helper functions
- Import SDK types directly (RegisteredCommand, ResourceLoader, Skill)
instead of casting through anonymous inline shapes
- Isolate the system prompt SDK escape hatch into a minimal boundary
helper with a named interface
Behavior is preserved — all 9 real e2e tests still pass.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix: regenerate package-lock.json with resolved/integrity fields
The previous lockfile regeneration ran under an offline/cached npm
context and stripped resolved URLs and integrity hashes from existing
entries, which would break npm ci and Nix builds.
* fix: update Nix npmDepsHash for regenerated lockfile
* chore(server): bump @mariozechner/pi-coding-agent to ^0.67.68
* fix: update lockfile signatures and Nix hash
* fix(app): surface agent setter RPC errors as toasts
setAgentMode/Model/ThinkingOption/Feature errors from the daemon were
only logged to console, so failures like missing API keys silently
vanished. Add toast.error fallbacks at all call sites and hoist
ToastProvider above SessionProvider so session context can use it.
---------
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
Co-authored-by: github-actions[bot] <github-actions[bot]@users.noreply.github.com>
Service hostnames now follow DNS convention with a project label so
independently opened projects no longer collide on routes like
`web.localhost`. Format becomes `{script}.{project}.localhost` for
default branches and `{script}.{branch}.{project}.localhost` for other
branches. The project slug comes from the GitHub repository name when a
remote exists, otherwise from the workspace directory basename.
Server changes:
- `buildScriptHostname` takes `{ projectSlug, branchName, scriptName }`
and slugifies all labels at the boundary; `untitled` fallback only
applies at hostname-label sites, never to shared `slugify`.
- `deriveProjectSlug(cwd)` and `parseGitHubRepoNameFromRemote` derive
the slug; GitHub URL parsing now requires `host === "github.com"` to
avoid embedded-path false matches.
- `ScriptRouteEntry` carries `projectSlug`; branch-rename handler reuses
it instead of re-reading git in hot paths.
- Status projection derives slug once per call and uses route hostname
when available; only falls back to building when no route exists.
- Spawn cleanup removes routes by `{ workspaceId, scriptName }` so
cleanup after a branch rename doesn't leave a stale entry behind.
App and shared message tests updated to the new fixture shape; service
URLs continue to be opaque server-provided strings.
Landing-page service URL examples updated; allowlist docs unchanged.
- worktree-core test scans the server source via a pure Node walker
instead of shelling out to `rg`, which is not installed on CI runners.
Invariant and scope are unchanged.
- settings-host-page specs target the new detail-header label + rename
affordance (settings-detail-header-title / host-page-label-edit-button)
since the label card moved into the shared header as part of the
modular settings refactor.
- settings-navigation compact tests assert the new two-tier Back
semantics: root Back exits settings, detail Back returns to /settings.
Introduce AgentStreamCoalescer to batch rapid-fire agent events before
broadcasting, reducing WebSocket chatter during tool-heavy turns. Wire
websocket-server runtime metrics so we can observe coalescing behavior
in tests. Work in progress.
Pull service env construction and port allocation out of worktree-bootstrap
into dedicated modules with their own tests, so the bootstrap path only
orchestrates. Adds peer collision detection and a port plan that is
computed once and reused across scripts.
Pass NO_BROWSER=true when probing ACP agents for models, modes, and
persisted sessions. Gemini CLI's ACP newSession handler calls refreshAuth
inline and will open a Google OAuth URL in the browser when no valid
cached token is present — every snapshot warm-up triggered this on
navigation. Real agent sessions still use their own launchContext.env,
so interactive auth continues to work when actually starting an agent.
Fan out agent close + terminal kill concurrently with Promise.allSettled
so one failure never leaves the worktree half-dead. Await real PTY exit
with SIGHUP→SIGKILL escalation before rm to avoid EBUSY races. Tolerate
missing repoRoot and retry directory removal with backoff. Purge stale
client workspace layout/tab state after successful archive.
The legacy list_provider_models_request handler was calling
providerRegistry[p].fetchModels() directly, so every client call spawned
the provider fresh (e.g. repeated copilot ENOENT errors). Old clients
still send this RPC.
Route it through ProviderSnapshotManager so simultaneous calls dedupe
through the in-flight warmUps map, cached entries (ready / error /
unavailable) are served without spawning, and at most one spawn occurs
per cwd per TTL window regardless of client count. Direct-fetch path
kept only as a fallback for the null-manager case (tests).
If a first archive attempt removed git's admin dir but failed to clean the
working tree (e.g. due to file churn during setup), retrying the archive
failed the ownership gate with "Worktree is not a Paseo-owned worktree"
because isPaseoOwnedWorktreeCwd required the worktree to still appear in
git worktree list.
- isPaseoOwnedWorktreeCwd now decides ownership by path shape
($PASEO_HOME/worktrees/<hash>/<slug>[/...]) and treats git as
best-effort for resolving repoRoot.
- deletePaseoWorktree tolerates a missing admin dir: it skips teardown
when the tree is absent, swallows a failing git worktree remove so
the rmSync fallback always runs, and follows up with git worktree
prune to clear stale admin refs.
- Relabel the post-creation setup failure log as
"Background worktree setup failed" since the worktree itself already
exists at that point.
Assistant markdown images with ~-prefixed paths hit ENOENT because
path.resolve doesn't expand ~, leaving paths like
workspaceRoot/~/rest-of-path that never resolve. Reuse the existing
resolvePathFromBase helper so tilde paths inside the workspace load
instead of getting stuck on the image spinner; the sandbox check still
rejects tilde paths that resolve outside the workspace.
Adds claude-opus-4-7 and claude-opus-4-7[1m] alongside the existing 4.6
entries (4.6 stays default). Opus 4.7 introduces a new "xhigh" effort
level between high and max; the option only surfaces on 4.7 models.
The SDK 0.2.71 effort union doesn't yet include xhigh, so the value is
cast at the assignment site.
Session.unarchiveAgentState(id) and the direct agentManager.unarchiveSnapshotByHandle
delegation were removed when main's 86bb5cc8 + a11c246b refactors moved unarchive
through the shared mcp-shared helper. Remove the now-dead test assertions so the
suite reflects the real ownership boundary.
