Make the repo's paseo skills self-contained — drop mentions of
personal skills (unslop, think-harder, e2e-playwright, fix-build,
rebase) that aren't checked in here.
The reconciler updated project.kind when a directory became a git repo
but never updated the workspace.kind underneath, leaving the sidebar in
a split-brain state: an expandable git project containing a workspace
with no leading icon (workspaceKind stayed "directory" instead of
flipping to "local_checkout" or "worktree").
* Fix relay E2EE reconnect races
* fix(relay): use WebSocket protocol ping for daemon control keepalive
The daemon's control socket pinged the relay every 10s with a JSON
`{type:"ping"}` message. Every JSON message wakes a hibernated Cloudflare
Durable Object via `webSocketMessage`, so an idle daemon was billing
relay DO CPU 24/7 just to echo `pong`.
Switch to `socket.ping()` (RFC 6455 protocol ping frames). Cloudflare's
runtime auto-responds at the edge without waking hibernated DOs, so an
idle daemon-to-relay connection now costs zero DO CPU.
Liveness still works: the `ws` library emits a `pong` event on the
daemon when the edge replies, which refreshes `controlLastSeenAt`.
`markControlReady` stays tied to DO-originated messages (sync,
connected, disconnected), so we still detect a broken DO at connect.
The DO keeps its JSON ping handler tagged `COMPAT(relay-json-ping)` so
old daemons (< v0.1.76) continue to work.
* test(relay): assert protocol-ping pongs arrive in e2e idle test
Adds a `relay_control_pong_received` debug log to the daemon's pong
handler, then asserts in the existing idle-relay e2e test that:
- at least one pong arrived during the 12s idle wait (proving the
Cloudflare edge auto-responds to RFC 6455 protocol pings)
- no stale terminations fired (proving the keepalive path is healthy)
If the daemon ever regresses to JSON-level keepalive pings (which wake
the hibernated relay Durable Object and incur CPU billing), this test
will catch it: the `pong` event would never fire because the DO's reply
would arrive via `message`, not as a protocol pong frame.
The test's capturing logger is bumped to `debug` level so the assertion
can read the new debug log; production daemons stay at `info`.
Confirmed locally against wrangler-hosted relay: 2 pongs / 0 stale per
12s window.
* test(relay): catch dual-ping regression via DO-side legacy ping log
The existing idle-test assertion catches "revert to JSON ping only" via
the missing pong event, but misses the dual-ping regression where
someone adds `socket.send(JSON.stringify({type:"ping"}))` alongside
`socket.ping()`. Both paths would work, the pong assertion still
passes, and DO CPU billing silently returns.
Add a console.log in the DO's handleControlKeepalive (which only runs
when a JSON ping reaches the DO), then assert in the idle e2e test
that no such log appeared during the 12s idle window.
Verified the assertion catches the regression: temporarily injected a
`socket.send(JSON.stringify({type:"ping"}))` next to socket.ping(),
test failed with `expected 1 to be 0` and the relay logged
`legacy_json_ping_received`. Removed the injection.
Log noise during the COMPAT window doubles as useful telemetry: it
counts how many old daemons are still pinging, which informs the
removal date for the COMPAT shim.
* fix(app): auto-set ref picker from a single attached PR
Attaching a PR in the new-workspace composer signals "I want this PR's
context", and almost always also means "base the worktree on this PR".
Until now the two paths were independent: the ref picker stayed on the
default branch, so the worktree was branched off main even though the
sidebar chip looked like a PR checkout.
Auto-promote the picker when exactly one PR is attached and the user
hasn't manually picked a ref. Manual picker selections are preserved;
auto-promoted ones are cleared when their backing attachment is removed.
* refactor(app): derive auto-promoted picker selection instead of syncing
The previous version drove `pickerSelection` from a useEffect that
inspected attachments and overwrote state. That meant the shared
`PickerSelection` type had to grow a `source: "manual" | "attachment"`
discriminator just so the effect could tell its own writes apart from
manual picks. Classic effect-as-derived-state.
Split the state instead. `manualPickerSelection` holds explicit picks;
`autoPickerItem` is a useMemo over composer attachments; the effective
`pickerSelection` is `manual ?? auto` via a small precedence helper. No
effect, no `source` flag, and `PickerSelection` stays a screen-local
shape since nothing outside needs it.
build runs on push to main and every PR including forks. Same model as
the test/lint/typecheck workflows: read-only token, no secrets injected,
runs scripts/update-nix.sh + nix build to validate. No commit, no push.
update-hash runs only on push to main. Mints the paseo-ai App token,
re-runs the update + build, and commits the refreshed hash via the
ruleset bypass. The default GITHUB_TOKEN no longer pushes anywhere.
Rename file to nix.yml since it's no longer just a build job.
scripts/update-nix.sh and scripts/fix-lockfile.mjs are executed from
the PR's checked-out tree. A fork PR can modify those scripts to run
arbitrary code on the runner. Same-repo PRs and push-to-main still
run; fork PRs skip the job entirely.
Closes#365.
GITHUB_TOKEN no longer pushes anything in this workflow. PR runs are
read-only, and main pushes go through the paseo-ai App token. The
workflow-level 'contents: read' is sufficient.
Addresses #365.
The hash-update commit only changes package-lock.json and
nix/npm-deps.hash. The parent commit already passed all 13 required
checks, and the file diff is content-only, so re-running the full
suite on every hash bump is wasted CI minutes.
Fork PRs cannot read repo secrets, so the App-token step failed with
'Input required and not supplied: app-id'. Gate the App-token step on
push-to-main and fall back to github.token for the checkout on PRs,
since PRs only need to validate the build.
OpenCode 1.14.42+ closes the /event SSE stream cleanly right after
server.connected, breaking the entire turn lifecycle: prompts get
queued and run, but Paseo never sees session.idle / message deltas /
tool calls / questions. Every turn either fails with a generic stream
EOF or hangs waiting for events that never arrive.
Switch the post-EOF recovery to the canonical messages REST endpoint
(via the upgraded SDK) and poll incrementally so tool calls and
clarifying questions surface live during the SSE gap. Bound the wait
with a completion cap and a separate liveness cap so silent rejections
fail fast instead of hanging until the cap. Cap session.abort similarly
so explicit cancels land within seconds.
The recovery path is gated on the for-await loop exiting without a
terminal event, so healthy turns never enter it - if upstream restores
SSE delivery, this code becomes dead and is mechanically removable via
the COMPAT(opencodeEofRecovery) and COMPAT(opencodeSlowAbort) tags.
Also tighten the type of projectSettingsRoute on WorktreeSetupCalloutPolicy
so router.navigate accepts it (was widened to string, broke app
typecheck on main).
Refs: getpaseo/paseo#861, anomalyco/opencode#26697,
anomalyco/opencode#26635
cancel-in-progress is gated to pull_request events so merge_group and
push runs always complete — only redundant CI from rapid PR pushes
gets cancelled.
Without merge_group on the workflow, PRs entering the merge queue would
fail because no required check ever reports a status against the
merge_group ref.
The matrix in CI / server-tests (ci.yml) covers what Server CI / test was
doing, on Linux and Windows instead of Linux only. Two workflows running
the same test suite on every server-touching PR is wasted runner time.
* test(codex): regression test for server-initiated JSON-RPC approval round-trip
Verified this test fails against the pre-#866 classifier shape from #758.
* test(codex): cover all server-initiated request methods in transport
* chore: add issue forms, PR template, contributor guidance
- bug report and feature request issue forms with structured fields for surface, version, provider, logs, screenshots
- issue config disables blank issues and routes casual questions to Discord
- single PR template with explicit verification section and AI-assisted contribution stance
- CONTRIBUTING.md: soften drive-by stance for objective bug fixes, codify expectations on AI-assisted reports and PRs (full logs over summaries, proof of testing over prose)
* chore: drop BDFL framing for one-person team
* chore: tighten CONTRIBUTING, drop duplication
The toolbar archive deferred its redirect until the server confirmed
the archive, so users briefly landed on the "workspace not found" gate
and were then sent to the main checkout's workspace route. Mirror the
sidebar pattern: synchronously router.replace to the project's new
workspace screen up front and let archive reconcile only update the
store.
Fork PRs (e.g. #845) run without OPENAI_API_KEY and without local
speech models, so global-setup.ts:420 threw and blew up the entire
Playwright job before any test could run.
Soften the gate: when neither path is available, warn and run with
dictation/voice disabled. Daemon already honors PASEO_DICTATION_ENABLED=0
and PASEO_VOICE_MODE_ENABLED=0. No spec currently exercises dictation,
so nothing else needs to change; future dictation tests should gate on
PASEO_DICTATION_ENABLED.
agent.integration "creates an autonomous live turn" — the prompt asked
Claude to use a `Task` tool that doesn't exist in the SDK (Paseo
synthesizes Task entries from sidechain notifications, but Claude can't
invoke it). Haiku non-deterministically refused or complied. Switch to
Bash with run_in_background, matching the deterministic
claude-autonomous-wake-simple.real.e2e variant — same autonomous-wake
mechanism is exercised.
schedule/service "defaults new-agent modeId to provider's unattended
mode" — afterEach raced AgentStorage's background snapshot persist,
producing intermittent ENOTEMPTY rmdirs. Drain pending writes with
agentStorage.flush() before rm.
Auto-namer now probes refs/heads/<name> and appends -2, -3, … (up to 50)
when the desired slug is taken, instead of failing with "branch already
exists" and leaving the worktree on its placeholder name.
Structured-generation waterfall replaces opencode/gpt-5-nano with
opencode/minimax-m2.5-free and opencode/nemotron-3-super-free so users
without nano configured fall through to free models they actually have.
Reload was a no-op for imported agents whose on-disk session was mutated
externally. The server preserved the in-memory timeline + historyPrimed
flag, so hydrate short-circuited and the epoch never changed. The client
also fetched without a cursor on reload, so the server returned
reset:false and the reducer dropped the new-epoch rows against a stale
cursor — the chat stayed stale until refocus.
Server: reloadAgentSession gains rehydrateFromDisk (used by the refresh
RPC). When set, durable + in-memory timeline are wiped and historyPrimed
is reset, so register mints a new epoch and hydrateTimelineFromProvider
re-streams the freshly read provider history. Voice-mode reloads keep
the existing preserve-timeline semantics.
Client: handleReloadAgent shows a sticky "Reloading agent…" toast,
sends the stale cursor with the fetch so the server returns reset:true,
and only flips to "Reloaded agent" after the response listener has
replaced the visible tail.
Filter persisted metadata-generation sessions out of the import list so
existing on-disk entries stop polluting it for users.
Rename the surface to "Import session" everywhere, swap the bespoke
filter pills for SegmentedControl with provider icons, drop the trailing
provider label on rows (icon is enough), tighten padding, and align the
row icon with the segmented control's left edge.
The GitHub picker showed "No results found." for any running agent after
the first message was sent. useAgentInputDraft owned a local cwd state
that clear("sent") wiped to "", so the composer read cwd: "" even though
the agent's real cwd was correct on the wire.
Drop cwd ownership from the hook. Composer takes cwd as a prop sourced
by the parent — agentState.cwd for running agents, composerState.workingDir
for drafts, sourceDirectory for setup flows. The hook now persists only
{text, attachments}. Legacy persisted drafts with a cwd field still
hydrate; tagged COMPAT(draft-cwd) for removal after 2026-11-09.
Also moves attachment pills back inside the composer's bordered surface
via a new attachmentSlot prop on MessageInput, and shrinks pill height
from 48 to 32 — the subagents-track tuck previously had attachments
floating between it and the input.
Adds one Playwright spec asserting the draft-create-agent path submits
the workspace cwd.
pickAttentionAgent now skips agents with a parentAgentId. The
workspace auto-focus path was opening subagents as tabs when they
needed attention, breaking the rule that subagents only become tabs
when the user explicitly clicks them.
Mohamed Boudra