Skip to content

RFC: Dynamic Task Dates ​

Task dates follow rules ("due 5 business days after Kickoff is completed") and are recalculated whenever something they depend on changes.

Owner: Ben TurnerCreated: 2026-10-06Status: Draft, in build behind dynamic_task_dates

Research, evidence and every product decision behind this plan: Dynamic Dates Research. How it works for the people using it: Automatic task dates (draft help article).

TL;DR ​

Rules, not dates

A date is calculated from its rule every time.

Nothing adds a shift to a saved date. The resolver is a pure function over a project snapshot, so the same inputs always give the same dates.

Trust

Every move is explained and undoable.

Original dates are kept, hand edits pin rather than break, a date waiting on unfinished work is labelled Estimated, and each move records its cause.

Aero

Aero can move dates on its own.

A new re-engagement action, governed by the existing autonomy settings, forecasts from history and pins the date it picks.

Motivation ​

17 named accounts (about $731K ARR) have asked for completion-driven dates, CRM or field anchors, or an end to re-dating by hand. Revco has turned dates off entirely and is a renewal risk. Today a playbook offset counts calendar days from project start or end, dependencies gate status only, and nothing moves when work finishes early or late.

Proposal ​

The rule ​

One sentence per date: [N] [business | calendar] days [before | after] [anchor].

AnchorMeaning
project_startThe project's planned start
project_go_liveThe project's go-live (end) date
task_start / task_dueAnother task's current start or due date, or this task's other date: a due date may follow its own start (a set length), a start date its own due date
task_completedAnother task's completion date; while it is open, its due date, and the result is Estimated
task_startedWhen another task was actually started (moved to In Progress or first answered); until then its start date (or due date), Estimated
task_answerA date answered in a date question of another task's form; until it is answered the date keeps its value (Waiting)
module_startThe start of the module this task belongs to
module_completedWhen every task in another module is finished or skipped; until then the latest of its due dates, Estimated
data_fieldA date-type data field, including CRM-synced fields

The editor picks the anchor in two steps: what to follow (the project, this task, this module, another task, another module, a date field), then which date about it. A rule may name its own task only through the other date, its own start (task_started, due dates only) or its own form (own_anchor_errors); never the same date or its own completion. Module and question anchors are stored as anchor_module_id / anchor_step_id + anchor_element_id, and follow their copies through reference_module_id / reference_step_id when a playbook becomes a project. A task being started or reset, and a form answer that a rule follows, recalculate the project through the same session hook that catches hand edits, so every writer of started_date and action_response is covered.

A task's start and due date can each carry several rules; the latest result wins. A module's start can carry rules too.

One rule schedules the whole task. When only one of a task's dates has rules, the other follows it, so nobody sets both by hand and a moving task never leaves its other date behind (or puts its due date before its start):

The rule-driven dateThe other date (no rules of its own)
Due date with a single rule after something happening (task_completed, task_started, module_completed)Starts when it happens: the same anchor, 0 days after
Any other due date ruleKeeps today's length before the due date, in business days when the due date's rules count business days
Start date with rulesDue date keeps today's length after the start

The decision is pure (implied_rule, task_length in resolver.py). A start that keeps a length is evaluated against the other date as it resolves in the same pass, so a task whose other date was empty settles in one recalculation (the idempotence fuzz test holds it to that). Nothing is implied for a date someone set by hand, a start date once the task has started, a date the other date already follows (task_start/task_due on its own task), or when there is no length to keep (no start date and a deadline-style due date); no date is invented. Implied dates explain themselves (implied: true on /projects/<id>/states, implied per task on the playbook schedule) and appear muted in the playbook list, tagged From due date / From start date in the Task dates drawer.

Behaviour (fixed, not configurable) ​

  • Only open tasks that have rules move. Tasks without rules and completed tasks never change.
  • A result before the project start is moved to the start and flagged. A module added mid-project never starts before the day it was added.
  • A date whose anchor is an unfinished task is Estimated: never shown as overdue, task status unchanged. Portal: Estimated · pending <vendor> unless the anchor task is customer-visible.
  • An edit by hand, an Aero move, or an undo of an Aero move pins the date, recording who: the rule is kept but not followed until someone re-links it. Any task can carry a pin, with or without a rule, so Aero knows which dates a person chose.
  • Completion days are counted in the vendor's time zone (or_vendors.primary_time_zone), so a task finished at 7pm in California counts that day.
  • When a rule-driven task or an Aero-pinned date lands past go-live, go-live moves with it, and it comes back in when that work finishes earlier, never before the committed go-live (the date set at creation, conversion, or the last time someone set go-live by hand or through the CRM). A date typed by hand does not move go-live; the existing "move the end date?" prompt still asks. original_go_live_dts is never touched, so "vs original" still shows the slip.
  • "Move dates earlier when work finishes early" is the one setting, per playbook and project, on by default. It applies only to moves caused by completions.
  • Rules that anchor to each other in a loop keep their current dates and are reported.

Data model ​

or_date_rules (one row per anchor), plus:

TableNew columns
or_tasksoriginal_due_date, original_start_date, due_date_pin, start_date_pin
or_modulesstart_date, added_to_project_at
or_tasks (state)due_date_estimated, start_date_estimated
or_projectscommitted_go_live_dts
or_projects, or_playbooksdynamic_dates (default off; new playbooks are created with it on), move_dates_earlier (default on)

Migration: migrations/versions/2026_10_06_6a711388bca6_add_dynamic_task_dates.py. The feature row ships disabled.

Build sequence ​

Everything ships behind dynamic_task_dates; the PR split is a delivery choice, not a product stage.

#SliceKey files
1Pure resolver, business-day maths, rule validation, model, migrationapp/api/date_rules/
2Recompute service: load a project snapshot, resolve, persist dates, go-live, originals, activity with causeapp/api/date_rules/services/
3Triggers: completion (four duplicated paths), reopen, date edits (pin), project date changes, data field saves, CRM syncservices_old.py, bulk_changes/events.py, data_fields_services.py, project/services.py
4Rules travel with tasks: playbook → project, library module insertion, playbook versions (via reference_task_id lineage)services_old.py:1991, project_create_service.py:1098, module_library/services.py:544
5Rule CRUD API, Convert, Chain this module, Re-linkapp/api/date_rules/routes.py
6One date control and editor (PrimeVue) in project and playbook lists and task panels; settings rowOTimeOffset call sites, TaskDetails.vue, PlaybookSettings.vue
7Overdue suppression for Estimated dates in the is_overdue hybrid and the eight raw comparisonstask/or_models.py:296, list in research doc
8Workflow add-modules anchor, automation addModule, manual addAddModulesConditionallyStep.vue, AddModuleAction.vue, StartFromModule.vue
9One digest email per cascade; due-date email recipientssend_task_due_date_changed_email, OREmailJob
10Insights "vs original" metrictask_performance_service.py, vendor analytics
11Aero recommend_date_adjustment: planner, replayer, hold rules, history read tool, date-moved triggersagent_autonomy/, onramp-agents/src/agents/re_engagement/

Alternatives considered ​

  • Shift saved dates by deltas (the 2022 cascade). Rejected: drift and dates appearing on tasks that never had one.
  • Behaviour modes per playbook (strict, flexible, none). Rejected: every competitor with modes draws the most complaints.
  • Dependency types and lag (FS/SS/FF). Rejected: an anchor sentence expresses the same thing without jargon.

Build status (uncommitted, 2026-10-06) ​

SliceState
Resolver, business days, validation, model, migrationBuilt; pure tests, Hypothesis idempotency, mutation-checked
Recompute service, originals, Estimated flag, activity with causeBuilt; database tests incl. tenancy, flag and transaction boundary
Triggers: completion, reopen, hand edits (pin), project dates, date data fields, creation, module insertionBuilt; hand-edit cascade covered by a session hook so every writer is caught
Rules copy with tasks (versions, creation, library, duplication)Built at the one dependency-copy choke point
Rule API, Convert (project, playbook), Chain this module, Re-linkBuilt; cross-tenant route test
Date control and editor, settings rows, Chain, ConvertBuilt (PrimeVue composites under components/composites/dateRules/)
Overdue ignores Estimated datesBuilt in the is_overdue hybrid and the nine raw comparisons
Workflow add-modules "Starts"Built: the project start or end date, project date fields (marked with their CRM source when the workflow maps them from the trigger), a playbook module or a module listed earlier in the step being completed, or a playbook task. Playbook modules follow the workflow's version counterparts (current_module_anchor); every module anchor resolves to its copy in the project through reference_module_id (module_anchor_in_project). Only earlier modules are offered, so two modules never follow each other
Date fields and CRM syncVerified on a real database through the call CRM inbound sync makes (test_a_crm_sync_into_a_date_field_moves_the_dates_that_follow_it); labels and activity name the field ("Due date moved because Contract signed date changed")
Portal "Estimated · pending <vendor>"Built in portalRenderer widgets
Insights "Finished by original due date"Built; shown only with the flag on
Aero move_date action, forecast, hold rules, Ask-mode cardBuilt; agent tool contract tests; an autonomous move past go-live carries go-live with it (dispatch test)
"Dates updated" emailBuilt: one email per project per action, to the project's internal members except the person who acted; see Email below

Verified end to end in a local clone (onramp_dynamic_dates): completing a task moved its dependents and logged why; a rule edit cascaded; Estimated dates rendered; go-live extended past a late chain; Aero's Move dates action and hold rules appear in settings.

Email ​

One action sends at most one "Dates updated in <project>" email per project, however many recalculations it caused. A bulk edit that completes ten tasks sends one. Each task appears once with where it started and where it ended up. A date that moved and came back is left out, and so is a task getting its first date (a newly added module).

  • Sent after the commit, from completion, reopen, bulk edits, data field changes, rule edits, Re-link, Chain this module, and module insertion. A failure is logged and never undoes the change.
  • When someone moves go-live by hand and the existing "Project end date updated" email goes out, that is the one email; the digest is skipped.
  • A hand edit to a task date sends today's due-date-changed email only. Any tasks that move because of it are logged in activity but not emailed, so the person editing is not sent a second email.
  • Not sent for project creation, Convert, or Aero moves (Aero reports through its own activity feed).

Which tasks Aero may move ​

One pure decision, aero_move_refusal (app/api/date_rules/forecast.py), decides it for the forecast list, the autonomous run and a person's Accept. Aero forecasts every overdue task, in converted and unconverted projects alike, except one that:

Refused becauseWhy
it is completed or skippednothing left to schedule
it is blocked by an unfinished task, or its date is an estimateit follows that task: Aero moves the blocker and this one moves with it
someone undid Aero's last move of this datefinal until a person changes the date
someone set the date by hand and it has not passed yeta person's date is theirs until it slips

The forecast date must also be in the future, within 120 days, and later than the current date. A slipped hand-set date is treated like any other overdue date and follows the autonomy mode. Admins who want those confirmed add the Conditional rule "Someone set this task's date by hand", which is available but not added by default.

The "passes go-live" safety rule now checks where go-live would land after the move and every date that follows it, not just the moved date.

Pressure test (2026-10-06, local clone) ​

A harness drove the real HTTP routes as the vendor admin against onramp_dynamic_dates. It covered the following: building a 9-task playbook (6 rules, 2 dependencies); 8 past projects as history; creation; completion, bulk completion, skip; hand edits; the go-live shift; Aero autonomous, Conditional and Ask paths; undo through the activity card; re-link. Final run: 69 checks pass, 0 fail, 26 observations. The harness lives in the session scratchpad and is not in the repo. The clone was backed up first as onramp_dynamic_dates_pre_pressure.

Fixed because of it (each now has a test that fails without the fix):

  1. Aero's move was recorded as a person's edit. Agent operations commit before they apply the change, which dropped the "Aero is writing this" marker. The marker now lives on the task itself.
  2. Undoing an Aero move never pinned the date as undone, so Aero could redo it.
  3. Editing a date Aero had set left it marked as Aero's, so it kept pushing go-live.
  4. Forecasts ran for vendors with the flag off.
  5. Moving go-live with "move task dates too" would have pinned every rule-driven task, which silently turns off every rule in the project.
  6. Completion days were counted in UTC, so dates landed a day late for evening completions in the Americas.
  7. An Aero move past go-live on a task with no rule did not move go-live.
  8. The go-live safety rule missed a move whose follow-on dates crossed go-live.
  9. Blocked and skipped tasks were offered to Aero.
  10. The older form-based project settings save changed go-live and start dates without recalculating anything (wired in; no automated test yet).

Every check in the final run ​

Run 212357 on 2026-10-06: 69 pass, 0 fail, 0 errors, 26 observations (recorded, not judged). Whole run: 9.4 s; slowest request 0.35 s (POST /projects/create_with_data).

Setup ​

ResultCheckExpectedObserved
PASSnew playbook defaults to dynamic datesTrueTrue
PASSplaybook rules saved through the route66
PASShistory tasks finished (8 projects x 9 tasks)7272

1. Creation from the playbook ​

ResultCheckExpectedObserved
PASSproject copies dynamic dates from the playbookTrueTrue
PASScommitted go-live set at creation2026-11-132026-11-13
PASSrules copied to the project66
PASSevery dated task has an original date00
PASSno task is pinned at creation00
OBSERVEtasks shown Estimated at creation-['Admin training', 'Customer sign-off', 'Data import', 'Integration testing']
PASSLaunch call due on go-live2026-11-132026-11-13

2. Completion cascade + forecast eligibility ​

ResultCheckExpectedObserved
OBSERVEemails after completing Kickoff call through the route-['[Dev] PT Live 212357: task dates updated']
PASSData import due 3 business days after today's completion (vendor's local day)2026-10-092026-10-09 (completed at 2026-10-07 02:24:04.622612 UTC; vendor zone America/Los_Angeles)
PASSafter backdating Kickoff call to 30 days ago, Data import moves to 3 business days after that2026-09-092026-09-09
PASSAdmin training is Estimated (waiting on Data import)TrueTrue
OBSERVEEstimated dates already in the past (never shown overdue)-[('Admin training', '2026-09-16'), ('Integration testing', '2026-09-30'), ('Customer sign-off', '2026-10-02')] (they hang off Data import's overdue due date)
OBSERVEtasks Aero forecasts-['Data import', 'Security review']
PASSoverdue rule-driven Data import is forecastTrueTrue
PASSoverdue task with no rule (Security review) is forecastTrueTrue (due 2026-09-26)
PASSblocked + estimated Admin training is not forecastFalseFalse
PASScompleted Kickoff call is not forecastFalseFalse
PASSData import forecast from same-task history('2026-10-08', 'same_task_history', 8)('2026-10-08', 'same_task_history', 8) (27 days late; past slips [-1, 0, 2, 4, 6, 9, 13, 21])
PASSSecurity review forecast from same-task history('2026-10-26', 'same_task_history')('2026-10-26', 'same_task_history') (10 days late; past slips [0, 0, 1, 2, 2, 3, 4, 30])
PASSa skipped task drops out of the forecastFalseFalse

3. Hand-set dates (option 2) + the optional rule ​

ResultCheckExpectedObserved
PASSa date typed in the slideout is pinned as a person'suseruser
PASSa hand-set date still ahead is not forecastFalseFalse
PASSonce the hand-set date has passed, Aero forecasts itTrueTrue
PASSConditional + 'set by hand' rule sends the move to Ask('active', 'asked', 'task_date_set_by_hand')('active', 'asked', 'task_date_set_by_hand')
PASSthe date did not move while held2026-10-032026-10-03
OBSERVEConditional with no rules enabled-holds everything (existing autonomy behaviour)
PASSAutonomous moves a slipped hand-set date('resolved', 'executed')('resolved', 'executed')
PASSit is now pinned as Aero's('aero', 2026-10-10)('aero', 2026-10-10)
PASSa person accepting an Aero card on a future hand-set date is refused('stale', 'Someone set this date by hand and it has not passed yet')('stale', 'Someone set this date by hand and it has not passed yet') (HTTP 200 {"data": {"status": "stale"}, "message": "Someone set this date by hand and it has not pas…
PASSand the date stays2026-10-162026-10-16
PASSrace: a person re-dates the task after Aero's card, before it runsrefusedrefused ({'status': 'active', 'outcome': 'refused', 'reason': 'Someone set this date by hand and it has not passed yet', 'rule': None})
PASSthe person's date wins2026-10-142026-10-14

4. Aero moves past go-live, undo, and edits after ​

ResultCheckExpectedObserved
OBSERVEgo-live right after creation (asked for 2026-10-09)-2026-10-09 (committed 2026-10-09 00:00:00)
OBSERVEgo-live before Aero (the estimated chain already passed the committed date)-2026-10-14
OBSERVEdates before Aero-{'Kickoff call': '2026-08-29', 'Data import': '2026-09-21', 'Admin training': '2026-09-28', 'Integration testing': '2026-10-12', 'Security review': '2026-09-21', 'Custom…
PASSData import is overdue and forecastTrueTrue
PASSAero's autonomous move runs('resolved', 'executed')('resolved', 'executed')
PASSData import moved and pinned as Aero's(2026-10-16, 'aero')(2026-10-16, 'aero')
OBSERVEdates after Aero-{'Kickoff call': '2026-08-29', 'Data import': '2026-10-16', 'Admin training': '2026-10-23', 'Integration testing': '2026-11-06', 'Security review': '2026-09-21', 'Custom…
PASSgo-live moved out past the Aero dateTrueTrue (go-live 2026-10-09 -> 2026-11-10; committed 2026-10-09 00:00:00)
PASSoriginal go-live kept for 'vs original'TrueTrue (2026-10-09 00:00:00)
PASSLaunch call follows the new go-live2026-11-102026-11-10
OBSERVEemails from Aero's move-[]
OBSERVElatest task activity-['Task "Data import" Due date changed from Sep 21 2026 to Oct 16 2026 by Admin OnRamp', 'Task "Launch call" Due date moved because the go-live date changed (Oct 14 → Nov…
OBSERVEpropose-undo-HTTP 200 uuid=ca7358bf-fc4a-4758-86f1-c38ff9eda5a0
OBSERVEexecute undo-HTTP 200 {"data": {"blast_radius": {"comments": 0, "projects": 1, "tasks": 1}, "executed_at": "2026-10-07T02:24:07.318702+00:00", "intent": "Undo: Move Data import to Oc…
PASSundo puts the date back2026-09-212026-09-21
PASSundo pins it as undoneundoundo
PASSgo-live comes back in to where it was before Aero2026-10-142026-10-14 (committed 2026-10-09)
PASSan undone date is not forecast againFalseFalse
PASSa new Aero card on the undone date is refusedrefusedrefused ({'status': 'active', 'outcome': 'refused', 'reason': "Someone undid Aero's move of this date", 'rule': None})
PASSa person editing an undone date takes it overuseruser

5. Person edits a date Aero set ​

ResultCheckExpectedObserved
OBSERVEstarting pin-aero
PASSa hand edit replaces Aero's pin with a person'suseruser

6. Go-live shift with 'move task dates too' ​

ResultCheckExpectedObserved
PASSKickoff call+14 days (legacy shift)2026-09-28 -> 2026-10-12
PASSData import3 business days after Kickoff call's new date (its rule)2026-10-01 -> 2026-10-15
PASSAdmin training5 business days after Data import's new date (its rule)2026-10-08 -> 2026-10-22
PASSIntegration testing10 business days after Admin training's new date (its rule)2026-10-22 -> 2026-11-05
PASSSecurity review+14 days (legacy shift)2026-10-21 -> 2026-11-04
PASSCustomer sign-off2 business days after Integration testing's new date (its rule)2026-10-26 -> 2026-11-09
PASSGo-live prep5 business days before the new go-live (go-live Sat Dec 19)2026-11-30 -> 2026-12-14
PASSLaunch call0 business days after the new go-live (go-live Sat Dec 19)2026-12-07 -> 2026-12-21
PASSHypercare check-in+14 days (legacy shift)2026-11-20 -> 2026-12-04
PASSthe shift pins nothing[][]
PASScommitted go-live follows a person's go-live change2026-12-192026-12-19

7. One email per action ​

ResultCheckExpectedObserved
OBSERVEbulk complete 3 tasks-HTTP 200
OBSERVEemails sent by the bulk completion-['[Dev] PT Bulk 212357: task dates updated']
PASSat most one 'task dates updated' email per recipientTrueTrue
PASSData import moved off the completion2026-10-092026-10-09

8. Flag off, legacy projects, extremes ​

ResultCheckExpectedObserved
OBSERVElegacy (not converted) project 34 'New Playbook' forecasts-{'Unnamed Task': ('2026-10-09', 'no_history', 1), 'New Playbook': ('2026-10-09', 'no_history', 14)}
OBSERVElegacy (not converted) project 33 'New Playbook' forecasts-{'Unnamed Task': ('2026-10-09', 'no_history', 2)}
OBSERVElegacy (not converted) project 4 'Acme — Migration' forecasts-{'Data cleanup sprint': ('2026-10-09', 'no_history', 1)}
PASSa hand edit in a legacy project is recorded toouseruser
PASSflag off: Aero forecasts nothing{}{}
OBSERVEproject whose go-live passed 150 days ago: forecasts-{'Kickoff call': ('2026-10-08', 'same_task_history', 258), 'Security review': ('2026-10-08', 'same_task_history', 235), 'Hypercare check-in': ('2026-10-08', 'same_task_h…
PASSKickoff call: forecast is a weekday within 120 daysTrueTrue (2026-10-08)
PASSSecurity review: forecast is a weekday within 120 daysTrueTrue (2026-10-08)
PASSHypercare check-in: forecast is a weekday within 120 daysTrueTrue (2026-10-08)
PASSGo-live prep: forecast is a weekday within 120 daysTrueTrue (2026-10-08)
PASSLaunch call: forecast is a weekday within 120 daysTrueTrue (2026-10-08)
OBSERVEa task with no playbook lineage-('2026-10-09', 'no_history', 0)
ResultCheckExpectedObserved
OBSERVESecurity review (no rule) pin before re-link-None
PASSre-link clears the pin and the rule takes the date back(None,)(None,) (HTTP 200; due now 2026-09-21 00:00:00)

10. A move whose cascade passes go-live (Conditional safety rule) ​

ResultCheckExpectedObserved
OBSERVEsetup-go-live 2026-10-26; Aero moves Data import to 2026-10-23, before go-live; Admin training (5 business days after) would land past it
PASSthe safety rule catches a cascade past go-live('active', 'asked', 'date_passes_go_live')('active', 'asked', 'date_passes_go_live') ({'status': 'active', 'outcome': 'asked', 'reason': 'rule_matched', 'rule': 'date_passes_go_live'})
PASSnothing moved while held2026-10-262026-10-26
OBSERVEcontrol-a move to 2026-10-06 (today): the chain after Data import still ends past go-live, so this is held too (estimates downstream hang off Data import's overdue date until it…
OBSERVEcontrol outcome-{'status': 'active', 'outcome': 'asked', 'reason': 'rule_matched', 'rule': 'date_passes_go_live'}

11. Aero moves a task with no rule past go-live ​

ResultCheckExpectedObserved
OBSERVEsetup-go-live 2026-10-11; Security review (no rule) due 2026-09-01
PASSAero's move runsexecutedexecuted
PASSgo-live follows Aero's move of a task with no rule2026-10-262026-10-26
PASSLaunch call (on go-live) follows2026-10-262026-10-26

Automated tests added for the pressure-test fixes ​

Each was checked to fail with its fix removed.

TestGuards
test_aero_moves_a_date_only_when_nobody_and_nothing_else_owns_it (11 rows)The full refusal table: done/skipped, blocked/estimated, undone, hand-set before and after it slips, past, too far, already moved
test_a_completion_counts_on_the_vendors_own_calendar_day (3 rows)7pm Pacific counts that day; London and an aware instant convert too
test_a_pinned_date_stays_put_and_only_aero_pins_move_go_live (+2 rows)An Aero move on a task without a rule moves go-live; a hand edit does not
test_the_go_live_rule_holds_a_move_that_stays_inside_go_live_but_pushes_what_follows_past_itThe Conditional go-live rule sees the cascade
test_undoing_an_autonomous_date_move_puts_the_date_back_and_stops_aero_moving_it_againUndo through the activity card restores the date, pins it undone, and the next card is refused
test_an_autonomous_date_move_past_go_live_is_pinned_as_aeros_and_carries_go_live_with_itAero's write is never mistaken for a person's, even though the operation commits midway
test_moving_go_live_with_its_tasks_shifts_them_without_pinning_any"Move task dates too" never switches rules off
test_only_dates_that_already_existed_are_reported_as_movedFirst dates on a new module are not emailed as moves
test_a_dependency_suggests_starting_when_the_blocker_is_done (6 rows) + test_a_dependency_suggestion_never_replaces_a_rule (3 rows)The start follows the blocker with today's gap, the due date keeps today's length after the start, and an existing rule is never replaced
test_a_task_date_follows_its_own_task_only_through_the_other_date (7 rows)A date may follow its own task's other date, its own start, or its own form, never itself or its own completion
test_one_rule_schedules_the_whole_task (10 rows) + test_a_due_date_pulled_in_before_its_start_brings_the_start_with_it + test_an_implied_date_says_which_rule_it_followsStart when it happens, keep the length (business days), invent nothing, leave hand-set and started starts alone, never override a date's own rule
test_a_task_whose_start_was_empty_settles_in_one_pass_off_the_weekend + the widened idempotence fuzz (hand-set starts, started tasks, task_started)One recalculation settles every schedule
test_a_due_date_that_follows_its_own_task_starting_waits_instead_of_feeding_on_itself, test_a_date_that_follows_an_unstarted_task_waits_for_that_tasks_due_date_to_resolveTwo "is started" bugs the widened fuzz found: a due date estimating from itself, and an undeclared dependency on the started task's due date
test_a_start_that_already_waits_on_the_blocker_through_its_due_date_is_not_offered_it_againNo redundant dependency suggestion
test_a_crm_sync_into_a_date_field_moves_the_dates_that_follow_itA CRM value synced into a project date field moves the tasks that follow it, and the label and activity name the field
test_a_module_a_workflow_adds_starts_from_what_its_start_rule_names (3 rows) + test_a_module_start_anchor_finds_its_module_in_the_project (6 rows) + test_a_workflow_module_anchor_follows_the_playbooks_current_version (3 rows)Workflow modules start from the project end date, a date field, or a module added earlier in the step, matched to the right copy
test_following_a_dependency_moves_the_whole_task_when_its_blocker_finishes_lateFollowing on a real database: a blocker finished late moves start and due together, and a hand-set start follows again
test_a_dependency_suggests_its_start_follow_the_blocker_and_ignores_another_vendors_taskThe suggestion reads real dates and drops a blocker from another vendor

Suites run after the fixes: date_rules 107, agent_autonomy 198 + 3, agent 1894, task 1960, project 1758, data_fields 412, custom_email_templates 819, utils/email 429, playbooks 1235, time_tracking 2046, core_insights 373. All pass; ruff clean.

Usability walkthrough (2026-10-06, local) ​

Walked through as an admin setting up the pressure-test playbook, and as a project owner on its projects.

Admin: would they know to give a task a dynamic date? No.

  • The playbook list's Due Date column still shows the old offset ("+5 days") for a task whose date actually comes from "3 business days after Kickoff call is completed". Nothing in the list says a task has a rule.
  • The only place to set a rule is a "Date rules" panel under Due Date in each task's slideout, one task at a time. Directly above it the slideout still shows "Due Date +5 days", so the admin sees two answers for one date and nothing says the rule wins.
  • Settings shows "Dynamic dates: On" as a badge with no link to where rules are set, next to "Exclude Weekends: Off" while the rules count business days.
  • On a read-only (published) playbook the rule controls are disabled but do not look it.

Owner or assignee: would they know what a date is based on? Only if they open the task.

  • In the slideout, yes: "Estimated · follows Data import", the rule sentence, and "Originally Sep 15".
  • In the list, a date set by a rule from a finished task, a date typed by hand, and a fixed date all look identical. Only Estimated and Pinned dates carry a chip, against the "indicators everywhere or nowhere" rule.
  • The Estimated chip pushes its date out of the Due Date column ("6 · Estimated").
  • An estimate already in the past (Sep 16, waiting on an overdue task) shows only a past date and a chip, with no hint that the task it waits on is late.
  • Activity showed engine moves as "None added": fixed (see the pressure-test fixes).

Built in response (2026-10-06)

The test for every screen: someone opening the app for the first time can read why a date is what it is, in words, next to the date.

  1. One date control everywhere, one editor. Every task date, in the project list, the playbook list and both task panels, is one control (TaskDateField): the date with what decides it in plain words beside it (in a list: Sep 9, 2026 · after Kickoff call, estimate · Data import is 27 days late in amber, set by hand, set date; in a task panel the full sentence under it). Clicking the date opens the same editor (TaskDateEditor) everywhere, beside the row or field: Follows other work (the rule as a two-line sentence, Also wait for… for more rules, a plain result line) or Set a date (an inline calendar in projects, which saves through the page's own date handling so the "move the end date?" prompt still asks; days from the project's start or end in playbooks). A date set by hand offers Follow the rule again under it and Follow again in the editor. A playbook date that comes from its task's other date is shown muted. Where automatic dates are off, the page's own date input is shown unchanged.
  2. No separate explanation chip. The words beside a date explain it, and the thing that opens an editor is the date itself, which was already editable, so the "indicators explain, never edit" rule still holds. The date columns widen for projects and playbooks using automatic dates.
  3. One place to see the schedule. Settings → Automatic dates → View task dates opens a drawer with every task's start and due date by module (Follows or Set date, plus the sentence), Chain this module, and Turn on for a playbook still on set dates.
  4. Rules offered from dependencies. A task that waits on another task and whose start date has no rule is offered, on its start date, "This task waits on Kickoff call. Start it 2 days after Kickoff call is done? Its due date stays 4 days after it starts.", in the task and at the top of the drawer (Follow all). Following stores one start rule (POST /date-rules/tasks/<id>/follow-dependencies): the start follows the blocker's completion with today's gap, and the due date, without a rule of its own, moves with the start. A late blocker moves the whole task instead of squeezing it. A blocker the start already waits on through its due date's rule is not offered again. The suggestion is a pure decision (follow_dependencies).
  5. Estimates say why. An estimate names the unfinished task at the root of its chain and how late it is; a late one turns amber.

Plain words throughout: Automatic dates (not "dynamic dates"), project end date (the product's name for planned_go_live_dts; "go-live" is gone from these screens, the activity log, the email and Aero's rule), set by hand (not "pinned"), Follow the rule again (not "re-link"), set date (not "fixed"). Exclude Weekends now says it applies to set dates only.

Still open: set dates in the playbook list still use the inline "+2 days" editor, which does not say what it counts from. The Insights health rules and the portal's "Go-live Jul 1" label predate this work and still say "go-live".

Answered (2026-10-06) ​

#QuestionAnswer
1Which dates may push go-live out?Rule-driven dates and Aero moves. A hand edit keeps today's prompt asking whether to move the end date.
2Should go-live come back in on its own?Yes, when the latest task date comes back in automatically, but never before the committed go-live.
3Convert: calendar or business days?Calendar days, so nothing moves on conversion.
4Chain this module: same day or keep gaps?Same day for now; revisit in local testing.
5Cascade emailOne email when dates change.
6Email when go-live moves itself?Covered by the same single email; revisit in testing.
—Hand-set dates and AeroA person's date is theirs until it slips; then Aero may forecast it. Optional Conditional rule "Someone set this task's date by hand", not on by default.
—Blocked tasksSkipped by Aero: they follow their blocker, which Aero moves instead.
—Projects not convertedAero forecasts dates in every project.
—Start dates (2026-10-07)Same treatment as due dates everywhere: list columns, date control, editor. A dependency suggests the start date.
16Estimates that have already passed (2026-10-07)An estimate never starts before today: while a task a date follows (completion, start, module) is still open, it is counted from the later of its planned date and the vendor's today (ProjectSchedule.today, _not_before_today). Recalculated on events only, never nightly. Resetting a finished task names it in the activity ("Due date moved because Kickoff call is no longer done").
—Editing dates (2026-10-07)Every date is editable where it is shown, project list included, through one editor with Follows other work / Set a date.
—Setting both dates (2026-10-07)Not required: one rule schedules the whole task (see The rule). Both dates can still have their own rules.
—Preview before a hand edit cascades (2026-10-07)Not built. A hand edit moves only the dates that follow it, every move is logged with its reason, and the existing "move the end date?" prompt still asks.
—Completion rules creating "blocked by" (2026-10-07)Not built. Rules and dependencies stay separate; a dependency instead suggests a start rule.
—Estimate wording (2026-10-07)Kept as built: "estimate · waiting on Kickoff call", plus how late it is. The portal keeps "Estimated · pending <vendor>".
—Who counts as editing a date by hand (2026-10-07)Dragging on the Gantt, bulk edits, automations, the Customer API and MCP are hand edits: the date is pinned "set by hand". A playbook update pushed to a live project is not: the project takes the new version's rules as if it had been created from it (follow_playbook_update__db_add_flush), its date writes are engine writes, and a date someone set by hand keeps their date. Undoing the update puts the rules back from a snapshot (project_rules_snapshot / restore_project_rules__db_add_flush).
21Playbook rule changes and rules changed in a project (2026-10-07)A change to a playbook task's date rules counts as a change to that task and is pushed like any other. A task whose rules were changed inside the project is held back as "edited in project", like other edited items. Rules are compared with task_rule_values; the update refreshes rules on every project and recalculates dates only where they are automatic.
—Turning the flag off or on (2026-10-07)A feature-gate reconciler (gate_service.py) runs after every admin write of the flag. Off: the organization's estimate markers are cleared, so a passed date reads as overdue again everywhere overdue is counted. On: each open project using automatic dates is worked out again in the background, with no emails. A flag changed outside the admin pages (raw SQL) skips the reconciler, as for every gated feature.
—Bulk actions and never-started tasks (2026-10-07)A bulk complete or bulk edit recalculates each project once, at the end, and sends one email per project (one_recalculation_per_project). Finishing a task nobody started no longer counts as starting it, so its completion respects Move dates earlier when work finishes early and its email is sent.
7What a customer reads beside an estimated date (2026-10-07)The date and what keeps it: "Oct 14 · if Sign the order form is done by Oct 9" when it waits on the customer's own task (never a day before today), else "depends on Acme's progress". Built on the portal's task rows.
8Start rule for modules an automation adds (2026-10-07)Built: the automation's Add module action has the same Starts picker as workflows. Modules added by hand still start the day they are added.
9Undoing a completion (2026-10-07)Resetting a finished task lists the dates that follow it and offers Reset the task and move those dates back or Reset the task, keep those dates where they are (kept dates read as set by hand).
10N/A tasks (2026-10-07)Not a task state customers or teams can choose; closed.
12Publishing a version that removes something a workflow's module starts from (2026-10-07)The publish dialog warns, naming the workflow, the module and what was removed; publishing is never blocked.
13Editing the "Dates updated" email (2026-10-07)The list of moved dates stays locked.
17Very late projects in Aero's forecast (2026-10-07)Not a valid case; closed.
18Organizations with no time zone (2026-10-07)Automatic dates count days in US Eastern, as integrations already assume. No screen sets an organization's time zone, so there is nowhere to prompt.
19Who the activity credits (2026-10-07)Aero, for everything it does on its own, including the dates that move because of it: "Aero · Autonomous · for Jane Smith".
22Two kinds of link on the timeline (2026-10-07)The internal timeline draws "Dates follow" links dashed beside "Blocked by" arrows, with a key; a task blocked by work its dates don't follow says so in its date editor.
24Dashboards on the day it is turned on (2026-10-07)Nothing changes when the flag turns on; numbers move only for projects that get rules, as they are created from a playbook with automatic dates or converted.
25Moving existing projects over (2026-10-07)Through playbook update propagation: turning on automatic dates in a playbook becomes a pushable project setting that converts each project. Wired here once this branch sits on the propagation PR.
20, 26Undo time of day; scale (2026-10-07)Not questions. A 400-task fully chained project recalculates in about half a second.
23, 14How many "Dates updated" emails people get (2026-10-08)Internal people get one Dates Summary at 8am in their own time zone (user's, else the organization's, else US Eastern), covering every project of theirs where dates moved since the last one; customers never get it. The person whose action caused a move is left out, unless Aero acted on its own, so Aero's moves are included. An email goes out right away only when a project's end date moves. Built on the deferred email queue: the engine logs moves to or_date_summary_items, keeps one queued job per person, and the queue builds the email just before it sends it. Admins can switch it off or customise its intro in Settings → Emails.
15Role recipients for the dates emailExpected: picking roles replaces the default audience (internal members minus the actor), as for every role-routed email.

Questions for Ben ​

Each question says who meets it, what they see today, and the choice that changes it.

11. Date rules on library modules. Parked for the follow-on module work: an admin editing a shared library module cannot see or set date rules there.

Housekeeping before any commit ​

  • The migration's down_revision must be repointed at main's current head when the branch is synced (main has since added a join-heads migration).
  • The email template migration (2026_10_06_d3a8c5e17f42) chains off this branch's migration; both move together when the head is repointed.
  • Tests whose assertions changed on purpose: the autonomy counters and policy payload (new move_date action), the portal frozen key set (due_date_estimated), the re-engagement tool count (6), and task fakes that now read due_date_estimated / original_due_date.
  • Legacy portal pages (pages/v2, pages/v3 chassis) still show due dates without the Estimated label; only portalRenderer widgets were updated.

Internal documentation — gated behind Cloudflare Access.