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
* fix(server): recover completed opencode turns after SSE EOF (#861)
* fix(server): wait for opencode completion after EOF
The original EOF recovery only checked OpenCode storage once after the SSE stream ended. That missed the real failure mode from #861: OpenCode can drop /event while the turn continues behind the scenes, then persist the assistant completion a little later.
Poll the persisted session for the active turn before failing EOF, while still requiring strong completion evidence and ignoring messages that predate the turn. This preserves failure behavior when there is no persisted completion.
Add behavioral coverage for delayed completion after EOF, partial streamed text plus persisted completion without duplication, stale old completions, and the no-evidence failure path.
---------
Co-authored-by: pluto <plutofog@proton.me>
Co-authored-by: Mohamed Boudra <boudra.moha@gmail.com>
* 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
* Patch production dependency advisories
* ci: harden dependency installs with npm ci --ignore-scripts
- Replace npm install with npm ci --ignore-scripts across all workflows
to enforce lockfile parity and block dependency postinstall scripts.
- Run npm run postinstall explicitly after install so our root patch
script still applies (the only legitimate postinstall use in this
repo); dep lifecycle scripts stay blocked.
- Add lockfile-lint and npm audit signatures to the lint job to catch
registry-host tampering and verify cryptographic signatures.
- Regenerate package-lock.json to match bumped package.json versions.
* ci: drop --ignore-scripts; rely on npm ci + lockfile-lint + audit signatures
Lavamoat allow-scripts only traverses root deps; it can't see workspace
deps like electron, esbuild, sharp. Without proper monorepo support, an
allowlist would be incomplete and electron's binary download breaks.
Keep the rest of the security stack:
- npm ci (strict lockfile parity)
- lockfile-lint (resolved-host check)
- npm audit signatures (cryptographic verification)
Real script-blocking would need pnpm 10+ migration. Tracking separately.
* ci(nix-build): use PR head SHA so fork checkouts work
The previous `ref: ${{ github.head_ref || github.ref }}` made
actions/checkout fetch a branch name that only exists on the fork,
not origin. Use the PR head SHA instead — origin mirrors PR commits
via refs/pull/N/head, so this works for fork PRs without changing
the push-to-main auto-commit behavior.
---------
Co-authored-by: Mohamed Boudra <boudra.moha@gmail.com>
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.
* feat(server): project-level prompts for metadata generation
Add `metadataGeneration` config to `paseo.json` so projects can customize the
four LLM-generated metadata strings — agent title, worktree branch name, commit
message, and pull request title+body. Each entry takes an `instructions` string
that gets injected as a `<user-instructions>` block between the default rules
and the load-bearing JSON format contract, with wrapper text saying user
instructions override defaults. Missing/empty/invalid config keeps every prompt
byte-identical to today.
* refactor(server): centralize project metadata prompt building
Three near-identical readers and four gate-then-wrap blocks collapsed
into one buildMetadataPrompt helper. wrapWithUserInstructions tightened
to require a non-empty string; its misleading empty fallback is gone.
* feat(app): edit metadata generation prompts in project settings
Adds four textareas (agent title, branch name, commit message, pull
request) under project settings, modeled on the existing setup/teardown
pattern. Round-trips through paseo.json preserving unknown sibling
fields at both the metadataGeneration and entry level.
Claude Code ignores --no-session-persistence outside --print mode, so the
SDK's persistSession=false is silently dropped in stream-json mode and the
metadata/branch-name generator sessions kept showing up as resumable. Sweep
the transcript on close when persistSession=false so internal generator runs
no longer leak into `claude` session listings.
* Fix chat everyone fanout
* Unify system-injected agent prompts
Chat mentions, schedule fires, and notifyOnFinish now share one shape:
prompt wrapped in <paseo-system>...</paseo-system>, no user-message
turn recorded, no auto-unarchive. Agents can recognize system-injected
context vs a real user turn.
- Move agent prompt orchestration out of mcp-shared.ts into a neutral
agent-prompt.ts module (mcp is an interface, not the home).
- Add `unarchive` and `recordUserMessage` flags to sendPromptToAgent.
- Add formatSystemNotificationPrompt helper.
- Schedule fires now identify themselves: "Schedule \"<name>\" fired
(id=..., run=...)." Plumb runId through runner so the paseo.schedule-run
label matches the actual run record.
- Replace chat-mention validateChatMentionFanout + duplicate
resolveChatMentionTargetAgentIds with single prepareChatMentionFanout
that owns expansion + cap + poster snapshot.
* Restrict provider import fan-out to enabled, non-derived allowlist
Move provider eligibility from a session-level post-filter to the agent
manager's fan-out so disabled providers are never spawned (fixes the
Gemini-disabled hang) and derived providers (like a custom Z.AI extending
Claude) don't duplicate their base provider's sessions.
Add an IMPORTABLE_PROVIDERS allowlist (claude, codex, opencode); ACP-based
providers are excluded from import discovery entirely.
* Filter Codex persisted threads by cwd before hydrating turns
Codex's listPersistedAgents previously called thread/read includeTurns on
every thread returned by thread/list, even when the import sheet was
scoped to one cwd. With many recent threads outside the scope, that fan
of reads dominated the import-sheet open latency.
Add an optional cwd hint to ListPersistedAgentsOptions, thread it through
the agent manager, and have Codex pre-filter the cheap thread/list rows by
cwd before scheduling the per-thread hydration. Widen the thread/list
window when filtering since most rows will be from other cwds. Other
providers ignore the hint.
* Fan out import session fetch per enabled provider
Sheet now reads the providers snapshot already warmed by the workspace
screen, filters to enabled importable providers (claude/codex/opencode),
and issues one fetchRecentProviderSessions request per provider via
useQueries with limit 15. Results are merged, deduped by handle, and
sorted by last activity.
Falls back to a single unfiltered request when the daemon does not
expose a provider snapshot (older servers). Per-provider failures show
inline so a single broken provider does not hide the rest.
* Add provider filter badges to import sheet
Show "All" plus one badge per enabled importable provider above the
session list when more than one is available. Single-select narrows the
already-merged list — no refetch. Each row now shows the provider's
icon next to its label so multiple providers stay visually distinct.
Selection resets when the sheet closes.
* Distinguish already-imported empty state in import sheet
Track how many descriptors were dropped because their handle was
already imported, surface the count back to the client as an optional
`filteredAlreadyImportedCount`, and switch the sheet copy to
"All recent sessions are already imported." when the merged list is
empty solely due to that filter.
* Cover already-imported provider session filtering
* List persisted OpenCode sessions for import
* Revalidate import sessions when sheet reopens
* Add agent persistSession plumbing
* Drop redundant visibility-transition refetch and test boilerplate
The import sheet's queries are gated on `enabled: visible`, and react-query
already refetches stale data when `enabled` flips on, so the manual
`wasVisibleRef` + `invalidateQueries` effect was redundant. Collapse the
two `selectedProvider` reset effects into one.
Drop `expect(...).toBeTruthy()` after `findByText` / `getByText` — those
helpers throw on miss, so the assertion was noise.
Drop `| null` from `buildOpenCodePersistedAgentDescriptor`'s return type
(it never returns null) and the dead `Boolean(descriptor)` filter at the
caller.
* Add provider-opaque session import path
Introduces fetchRecentProviderSessions on the daemon client plus an
importAgent overload accepting providerId/providerHandleId, and a
projection helper that renders persisted descriptors as provider-opaque
recent-session payloads with prompt previews.
* Unslop the import-session feature
Centralizes the importable-providers list, normalizes import_agent_request
shape at one boundary, and removes the per-provider fan-out fallback in
the import sheet. Old daemons without features.providersSnapshot now show
"Update host" instead of being served a degraded path.
Replaces internal-helper unit tests on the OpenCode and Codex providers
with boundary tests through listPersistedAgents; makes storageRoot
injectable on the OpenCode provider so the boundary test can target a
temp directory.
Splits the CLAUDE.md back-compat rule into two contracts: protocol stays
backward-compatible (always); features may require new daemons (no
fallback paths). Tags the two new compat shims with COMPAT(name) and
target removal dates.
* Extract importable-session listing into its own module
Move listImportableProviderSessions out of session.ts and drop the
unused generic listPersistedAgents from AgentManager.
