Add scheduled recurring agent runs

[0x92]

Description

Let a project start agent sessions on a schedule: a fixed prompt, at a fixed time, in a fresh worktree.

Pane already knows how to create a session from a prompt — this makes that repeatable without a person present. Nightly bug sweeps, a weekly dependency check, a "review yesterday's diffs" pass every morning: the work that is worth doing regularly but never worth remembering.

A new Scheduled runs entry in the project's menu manages them.

What it does

• Three schedule shapes rather than cron syntax: every N minutes, daily at HH:MM, weekly on a weekday at HH:MM. Each one can be stated in a sentence, which is what the dialog shows back to you (Every day at 03:00 — next run in 4h 12m). Cron is more expressive than anyone actually needs here and impossible to render as a sentence.
• Each run creates a real session, exactly as the create dialog would: prompt, agent (claude, or none for a plain terminal) and an optional worktree/branch template.
• Per-run state is visible: last run, its outcome, the error if it failed, and a link to the session it created.
• Enable/disable without deleting, and Run now — which starts one immediately without moving the cadence, so testing a schedule at 15:00 does not shift the nightly run.
• Missed runs are skipped, not caught up. Waking a laptop after a weekend must not start three days of nightly sweeps at once; a run more than 15 minutes late is recorded as skipped and the schedule moves on.
• Deleting a project deletes its schedules (ON DELETE CASCADE).

How it works

• scheduleCalculator.ts is pure and holds every decision about when: computeNextRun, isDue, isMissed. Time zones, daylight saving and "the next weekly run when today already passed" are the parts that break silently, so they are unit-tested directly rather than through the scheduler.
• ScheduleManager ticks every 30s and asks the calculator; it talks to storage through a small ScheduleStore interface, so its behaviour is tested without a database.
• Sessions are created through the existing task queue with createSessionAndWait, so lastSessionId is the real session id rather than a queue job id — the difference only shows up when you click the link.
• Persistence is one table, scheduled_runs, plus an index on (enabled, next_run_at_ms) so the tick is a single indexed lookup.
• Channels are daemon-owned (schedules:*): the schedule belongs to the machine that runs the agents.

Type of Change

• New feature (non-breaking change which adds functionality)

Checklist

• I have read the CONTRIBUTING.md guidelines
• My code follows the code style of this project
• I have performed a self-review of my own code
• I have commented my code, particularly in hard-to-understand areas
• My changes generate no new warnings
• I have added tests that prove my fix is effective or that my feature works
• New and existing unit tests pass locally with my changes
• I have run pnpm typecheck and pnpm lint locally
• I have tested the Electron app locally with pnpm electron-dev

Critical Areas Modified

• State management/IPC events

Additional Notes

Schema note: schema.sql is split on the statement separator at startup, so a ; inside a comment breaks the database. The new block is written accordingly.

Tested in the running app, not only in unit tests: a schedule was created, fired on its own tick, produced a real session in a fresh worktree, recorded its outcome, and was then disabled and deleted. Missed-run handling was exercised by moving a schedule's due time into the past.

Automated QA

Status: Passed on d14e9878 with synthetic project Scheduled QA.

• Playwright covered empty state, weekly form entry, save, pause/resume, Run now, persisted success state, and opening the resulting session.
• Node 22 gates passed: root typecheck, full lint, and 65 focused main-process tests.
• Native Electron dev launch succeeded with isolated data at /tmp/pane-pr386-qa. macOS accessibility automation was unavailable, so renderer interaction used the repository Playwright Electron API mock.

Empty state	Weekly schedule form	Run complete and paused

Remaining human check: allow one real scheduled run to fire in a fresh worktree with the preferred installed agent.

[parsakhaz]

Verdict: Request changes - four correctness defects violate the scheduled-run behavior stated in the PR.
Counts: Must Fix: 4 (security: 0) · Should Fix: 1 · pass 1/3

Must Fix

• MF-1 - Startup does not immediately process a run that is inside the 15-minute grace window · main/src/services/scheduleManager.ts:55-72 · run the first tick immediately after reconciliation · violates the stated missed-run grace behavior
  • Evidence: start() only installs a 30-second interval. The startup test helper manually calls manager.tick() at main/src/services/scheduleManager.start.test.ts:63-68, so it does not exercise the actual launch sequence.
  • Failure scenario: A run is 14m50s late when Pane opens. It is still eligible at reconciliation, but the first interval fires 30 seconds later and records it as skipped.
• MF-2 - In-flight execution is not coordinated with Run now, delete, disable, or edits · main/src/services/scheduleManager.ts:137-159 and main/src/services/scheduleManager.ts:209-240 · serialize by schedule id and merge execution results into the latest stored row · violates Run now starting one run and schedule mutation semantics
  • Evidence: runNow() bypasses the tick-level running guard, while execute() awaits session creation and later upserts the stale object it loaded before the await.
  • Failure scenario: A due tick and Run now can create two sessions. Deleting or disabling a schedule while its session is starting can later resurrect or re-enable it when execute() upserts its stale copy.
• MF-3 - The daemon boundary accepts malformed schedule variants that can never run · main/src/ipc/schedule.ts:18-34 · validate kind, toolType, integer interval/weekday, project id, and parsed clock time · violates the three supported schedule shapes
  • Evidence: the regex accepts 99:99, unknown kind values fall through as weekly, and runtime IPC input is trusted as ScheduledRunInput. computeNextRun() then returns null for invalid clock values.
  • Failure scenario: A remote or renderer caller saves an enabled schedule successfully, but it has no next run and silently never starts.
• MF-4 - The UI omits the promised link to the session created by the last successful run · frontend/src/components/schedule/ScheduledRunsDialog.tsx:167-172 · render lastSessionId as an action that closes the dialog and opens that session · violates the PR description under per-run state
  • Evidence: the row prints only time, status, and error even though lastSessionId is returned and persisted.
  • Failure scenario: A user sees a successful scheduled run but cannot navigate from the schedule to the session it created.

Should Fix

• SF-1 - Rebase residue adds unrelated and inconsistent declarations · frontend/src/types/electron.d.ts:39 and main/src/preload.ts:135 · remove the unused GitDiffResult import and unrelated pr: prefix
  • Evidence: GitDiffResult has no references in electron.d.ts, and pr: is added only to the preload copy, not shared/types/daemon.ts.

Praise

• The pure schedule calculator cleanly covers DST and cadence arithmetic.
• SQLite uses bound parameters and a foreign-key cascade, with no injection issue found.
• The focused suite passes: 4 files, 54 tests.

[parsakhaz]

Verdict: Approve - all four Must-Fix findings are resolved on d14e987.
Counts: Must Fix: 0 (security: 0) · Should Fix: 0 · pass 2/3

Must Fix

• MF-1 fixed: startup performs an immediate tick, including the grace-boundary regression.
• MF-2 fixed: execution locks by schedule id and merges results into the latest stored row.
• MF-3 fixed: daemon input uses the shared boundary decoder plus domain validation.
• MF-4 fixed: the dialog opens the last created session, covered by Playwright.

Praise

• Root typecheck and full lint pass under Node 22.
• Six focused test files pass with 65 tests.
• The scheduled-run Playwright journey passes.

[parsakhaz]

Review, simplify, refactor complete

Rebased onto current main and pushed final head d14e9878.

REVIEW

• Posted a request-changes review with four correctness findings.
• Fixed startup grace-window handling with an immediate tick.
• Prevented concurrent duplicate runs and stale execution writes from resurrecting or re-enabling schedules.
• Added shared boundary decoding and domain validation for schedule input.
• Added the missing Open last session action.
• Commit: e1dd708e fix(review): harden scheduled run execution
• Re-review posted with approval after all findings were verified.

SIMPLIFY

• Replaced conditional spreads and unchecked casts with explicit typed assignments and domain guards.
• Moved IPC shape decoding to the shared boundary schema.
• Removed dead schedule formatting code, dead exports, stale rebase imports, and redundant schedule clones.
• Blocking lint improved from 28 PR-introduced errors before the pass to clean.
• Commit: 5aa800f5 refactor(simplify): streamline scheduled runs

REFACTOR

• Bounded the indexed repository query to schedules due at the current tick instead of loading every enabled future schedule every 30 seconds.
• Added a repository contract test.
• Commit: 87c0b031 refactor(schedules): query only due runs
• Added durable renderer QA coverage for empty, create, save, pause/resume, Run now, and result-session navigation.
• Commit: d14e9878 test(schedules): cover scheduled run dialog

Verification

• PATH=/opt/homebrew/opt/node@22/bin:$PATH pnpm typecheck: passed
• PATH=/opt/homebrew/opt/node@22/bin:$PATH pnpm lint: passed, including Oxlint, ESLint, boundary conformance, advisory scan, and Knip
• Focused Vitest suite: 6 files, 65 tests passed
• Playwright scheduled-run journey: 1 passed
• Isolated Electron dev launch: passed with PANE_DIR=/tmp/pane-pr386-qa
• Screenshot bytes were uploaded to the existing pr-assets release and verified by direct download and SHA-256. The PR description contains the gallery.

Follow-ups

• Split the 366-line scheduled-runs dialog into orchestration, row, and form units after adding component-level interaction tests.
• Decide and document whether multiple schedules due together should be enqueued concurrently before changing current sequential execution.
• Add a native SQLite integration suite for row mapping, upsert behavior, and project cascade deletion.

Left for parsa

• Let one real scheduled run fire with the preferred installed agent and confirm it creates a fresh worktree and opens from the saved session link.
• Merge when satisfied. No merge was performed here.