Routines and pipelines
Routines and pipelines turn repeatable intent into controlled agent work. A routine starts work from time, an event, or an agent heartbeat and records the result. A pipeline gives that work a path through cases, outputs, conversations, and decisions. Together they connect an initial trigger to a durable work record and a result that people can inspect.
System model
Paperclip uses a routine for reusable agent work and a pipeline for repeatable case flow. A routine run creates or reuses an execution task. A pipeline case keeps the work, its context, and its outputs together as the case moves through its process.
The main relationships are:
| Concept | Purpose | Downstream result |
|---|---|---|
| Routine | Describe repeatable agent work | A tracked run and execution task |
| Trigger | Start or prompt a routine | A scheduled, event-driven, direct, or heartbeat-driven action |
| Pipeline | Define a path for one kind of case | A durable case flow |
| Case | Hold one item in a pipeline | State, context, links, and outputs |
| Case output | Collect work made by linked tasks | Documents, work products, and attachments |
| Conversation | Keep case discussion with its work | Context for people and later runs |
See Automation and control for the wider control model. See Decisions, approvals, and attention for the human gates that can stop or release work. See CLI and MCP interfaces for the shared operator interfaces.
Routines
Routine definition
A routine is a reusable unit of agent work with instructions, an agent owner, a responsible person, an optional project and parent task, input variables, environment bindings, and run policies. Its active or paused state controls whether automation may start. An archived routine is a terminal state and cannot run again. The routine description is also an editable document, so people can review instructions without losing the routine's execution identity. server/…/routines.ts• L2473-2567
Routine titles and descriptions can contain {{placeholder}} variables. Paperclip keeps variable definitions aligned with those placeholders and validates values by type. A scheduled routine must have defaults for its required variables because no person is present to supply them at the scheduled time. Secrets are bound to the routine environment and apply to the tasks that it creates.
Changes to the executable routine definition or its triggers create a revision that captures the routine and its active triggers. This gives operators a stable record of which configuration produced a run and lets them restore an earlier definition. A folder-only placement change does not change the executable definition. The description has its own versioned document representation, so instruction changes retain an edit history for review. server/…/routines.ts• L980-1039 server/…/routines.ts• L2640-2785
References: server/…/routines.ts• L2473-2567 server/…/routines.ts• L980-1039 server/…/routines.ts• L2640-2785
Routine suggestions
A routine suggestion turns a completed, repeatable task into an optional schedule. The assigned agent can offer the suggestion once, and a person can accept it or create a routine from the task with a chosen cadence and time zone. The new routine keeps the task's agent and project and can be filed under the responsibilities that contain the task. server/…/routine-suggestions.ts• L131-148 server/…/routine-suggestions.ts• L452-611
The system blocks suggestions that would create loops or expose private work. It does not offer routines for cancelled work, work already created by a routine, onboarding work, responsibility status or dashboard conversations, or other excluded automation conversations. A person-created routine also cannot come from a private task because routines are visible to the company. server/…/routine-suggestions.ts• L176-268
The responsibilities experiment controls this feature. If creation succeeds, the routine and its trigger are the durable result; filing links are useful placement metadata and do not undo the routine when a filing step fails. server/…/routine-suggestions.ts• L150-153 server/…/routine-suggestions.ts• L294-365
References: server/…/routine-suggestions.ts• L131-148 server/…/routine-suggestions.ts• L150-268 server/…/routine-suggestions.ts• L294-365 server/…/routine-suggestions.ts• L452-611
Trigger types
A trigger is the entry signal for a routine. One routine can have several triggers, so the same work can run on a clock, from an external event, through a direct request, or as a standing check on an agent heartbeat. Webhook triggers also support a pending setup state for safe connection checks before they can dispatch work. server/…/routines.ts• L2788-2860 server/…/routines.ts• L2896-2963 server/…/routine-standing-checks.ts• L15-24
| Trigger | Role | Configuration |
|---|---|---|
| Schedule | Start work at recurring times | Five-field cron expression and IANA time zone |
| Webhook | Start work from an external system | Public trigger URL, secret, signing mode, and optional pending setup |
| API | Start work on demand | No schedule data; fired through the run API |
| Heartbeat | Prompt an assigned agent to inspect a standing responsibility | Active routine, enabled heartbeat trigger, and an agent with its own timer wake |
The server calculates the next schedule time in the trigger time zone. Webhook requests can carry variables and payload data. Authentication checks the configured signature or bearer secret and can reject old or repeated signed requests. A pending webhook delivery records connection state without creating a run, task, or agent wakeup. A receipt also stops a setup test from starting work after activation. When testing ends, the last real delivery is restored instead of leaving the test result as the live delivery. server/…/routines.ts• L2862-3044 server/…/routines.ts• L3269-3349 server/…/routine-trigger-update-plan.ts• L10-44
A heartbeat trigger is different from the other trigger types. It does not start a run by itself. Each eligible timer wake presents the agent with its active standing checks, and the agent reports whether it acted or found nothing to do. Each report becomes a completed routine run, so the routine history shows the check without creating a task when no work is needed. server/…/routine-standing-checks.ts• L15-24 server/…/routine-standing-checks.ts• L117-149
References: server/…/routines.ts• L2788-2860 server/…/routines.ts• L2862-3044 server/…/routines.ts• L3269-3349 server/…/routine-trigger-update-plan.ts• L10-44 server/…/routine-standing-checks.ts• L15-24 server/…/routine-standing-checks.ts• L117-149
Standing checks
A standing check is a heartbeat-based routine that gives an agent a repeated responsibility instead of a repeated task. It runs only for an active routine with an enabled heartbeat trigger assigned to an agent that has an enabled timer heartbeat and can wake without a task. server/…/routine-standing-checks.ts• L26-39 server/…/routine-standing-checks.ts• L117-144
On each wake, the agent sees the check instructions and the last reported result. It performs the check before other wake work, creates or updates a task only when the finding needs work, and reports once for each check. The result records an outcome of acted or nothing_needed, an optional note, and an optional linked task. server/…/routine-standing-checks.ts• L218-245
Standing checks are experimental and require the responsibilities setting. Paused, removed, or unsupported checks do not accept a report. The last result is visible with the check and can include the linked task when the reader has access to it. server/…/routines.ts• L545-557 server/…/routine-standing-checks.ts• L56-100 server/…/routine-standing-checks.ts• L151-210
References: server/…/routines.ts• L545-557 server/…/routine-standing-checks.ts• L26-39 server/…/routine-standing-checks.ts• L56-100 server/…/routine-standing-checks.ts• L117-210 server/…/routine-standing-checks.ts• L218-252
Scheduling policies
Scheduling policies decide how Paperclip behaves when a new trigger overlaps existing work or when the scheduler returns after missed ticks. They belong to the routine, so all entry paths share the overlap rules while only schedule triggers use catch-up and activity gating.
| Policy | Meaning |
|---|---|
coalesce_if_active | Attach the new run to the live execution instead of creating duplicate work |
skip_if_active | Record the trigger but discard the overlapping run |
always_enqueue | Create work for every trigger occurrence |
skip_missed | Do not replay schedule times missed during downtime or a pause |
enqueue_missed_with_cap | Replay missed schedule work within the system cap |
require_external_activity | For schedule triggers, run only after qualifying outside activity |
The scheduler claims and advances each due tick before it checks pause, worktree, or activity suppression. This records suppressed work and prevents a later setting change from replaying an old tick. Catch-up is bounded. Archived or inactive triggers are not eligible for scheduler dispatch. server/…/routines.ts• L1501-1613 server/…/routines.ts• L3186-3305
Schedule calculation uses the trigger's IANA time zone. It supports sparse schedules without making a long search block unrelated scheduler work. server/…/routines.ts• L185-301
An activity gate can inspect the whole company or only the routine's project. It ignores the routine's own bookkeeping and read-only activity. Manual, API, and webhook fires do not wait for this gate because they are direct requests to run.
References: server/…/routines.ts• L1291-1416 server/…/routines.ts• L1501-1613 server/…/routines.ts• L3186-3305 server/…/routines.ts• L185-301
Routine run lifecycle
When a routine fires, Paperclip resolves its variables and execution context, records a run against the current routine revision, and creates an execution task for the assigned agent. The normal heartbeat flow then owns the work. A run can point to the created task, point to an existing task when work is coalesced, or finish as skipped or failed. server/…/routines.ts• L1784-2156
The dispatch record keeps the resolved inputs, routine revision, environment, workspace choice, source, payload, and result. Idempotency keys protect repeated requests from creating another run. The routine detail view exposes recent runs, the linked execution task, live work, and activity.
When the linked execution task changes state, Paperclip synchronizes the routine run. A completed task clears a transient blocked or cancelled result. A blocked or cancelled task records a transient failure that can be repaired without losing the run's history. server/…/routines.ts• L2044-2048
Issue and run changes also publish scoped routine-status work. A status worker finds unresolved run and execution-task pairs in bounded pages, then settles each pair with short database lock and statement limits. This gives the routine run a repair path when a status change does not complete the normal dispatch flow. server/…/routine-execution-status-notifications.ts• L3-8 server/…/routine-execution-status-work.ts• L7-26 server/…/routine-run-status.ts• L150-201
References: server/…/routines.ts• L1784-2156 server/…/routines.ts• L2862-2891 server/…/routines.ts• L3099-3173 server/…/routines.ts• L3186-3305 server/…/routine-execution-status-notifications.ts• L3-8 server/…/routine-execution-status-work.ts• L7-26 server/…/routine-run-status.ts• L40-201
Pipeline cases
A pipeline case is a durable work item that carries its own state, context, linked work, and results. The case surfaces below let later work and people understand what happened without depending on one transient execution task.
Case outputs
Case outputs are the durable work products reachable through a case's active task links. The output view gathers non-system documents, linked work products, and attachments. It reports each item with its source task, source role, trust information, and path back to the source. server/…/pipeline-case-outputs.ts• L266-518
Output order favors likely deliverable documents before other work products and attachments. Context summaries include a small number of bounded excerpts and provide fetch hints for full content. Low-trust content is redacted or marked as untrusted so an output can inform later work without becoming an instruction by accident.
References: server/…/pipeline-case-outputs.ts• L31-73 server/…/pipeline-case-outputs.ts• L146-218 server/…/pipeline-case-outputs.ts• L266-518 packages/…/pipeline.ts• L204-325
Conversation context
A case conversation keeps discussion attached to the case while its producing work changes. Paperclip chooses an active conversation source from an explicit conversation link, a recent producer update, the case's creation producer, automation work, or other linked work. A child case can inherit a parent producer as an inactive source when it has no local conversation link. server/…/pipeline-conversation-context.ts• L82-147
The case body is a versioned document. Opening a conversation links that body to the conversation task. Later edits update the document instead of rewriting the discussion task description. The context includes bounded body text, recent comments, linked runs, and output summaries.
Paperclip labels case fields, document text, comments, annotations, and output excerpts as untrusted content. It bounds the amount of content passed into context and provides explicit paths when an agent needs the full source. server/…/pipeline-conversation-context.ts• L249-327
References: server/…/pipeline-conversation-context.ts• L82-147 server/…/pipeline-conversation-context.ts• L249-327 packages/…/pipeline.ts• L204-325
Run continuation and recovery
Continuation keeps work attached to the same case or conversation when a run is still active or when a completed run leaves a clear next action. A new routine trigger can coalesce with the active execution. A conversation comment sent during a live run can join that run as queued input instead of starting competing work.
Bounded continuation
Run-liveness continuation handles a run that ends without concrete progress. It checks that the task and agent remain valid, the task is still assigned and active, no execution policy or hard budget stop blocks it, and no same wake already exists. It uses a bounded number of attempts with an idempotency key and sends the continuation through the normal model path. When the limit is reached, it records a notice that asks a person or manager to inspect the task. server/…/heartbeat.ts• L12451-12589
Max-turn exhaustion has a separate bounded continuation policy. Heartbeat can queue a delayed retry when the agent policy allows it. Provider and workspace failures can also delay the next attempt. Every path stops at its attempt cap or when an execution gate rejects the retry. server/…/heartbeat.ts• L14584-15399 server/…/heartbeat.ts• L15503-15644
When a successful agent run leaves an assigned task in progress, recovery requires a real disposition. The agent must finish, request review, record a blocker, delegate follow-up work, or set an explicit continuation. This prevents a report or comment from looking complete while no next action is recorded. server/…/successful-run-handoff.ts• L411-544
References: server/…/heartbeat.ts• L12451-12589 server/…/heartbeat.ts• L14584-15399 server/…/heartbeat.ts• L15503-15644 server/…/successful-run-handoff.ts• L411-544 server/…/heartbeat.ts• L12890-12994
Management flows
CLI and API management
The API is the shared control surface for routines, pipelines, cases, documents, outputs, conversations, and retries. Routine routes separate board management from agent self-management. Board actors manage company routines subject to task-assignment access. Agents can manage routines assigned to themselves and routines for which the person directing their run is responsible. Assignment, activation, and schedule changes still require the matching assignment authority. server/…/routines.ts• L107-169
The routes support routine changes, revision restore, trigger setup, secret rotation, manual runs, run history, description annotations, and public webhook delivery. Writes validate their input, record activity, and create routine revision records when configuration changes. The API can also return configuration warnings when protected environment keys need operator review. A heartbeat trigger can be added only when the responsibilities experiment is enabled. server/…/routines.ts• L214-251 server/…/routines.ts• L270-535 server/…/routines.ts• L545-716
The CLI maps common lifecycle actions to that API and supports human-readable output and JSON for scripts. cli/…/routine-api.ts• L21-95
The main operator flow is:
- Create or update a routine and assign its responsible agent.
- Add a schedule, event trigger, or standing check and choose the policies that apply.
- Start a run directly, wait for a trigger, or let an assigned agent report a heartbeat check.
- Inspect the run, linked task, case context, outputs, and conversation.
- Retry or continue work only when the next action and its limits are clear.
Useful command groups include paperclipai routine for direct routine actions, paperclipai pipelines for pipeline and case operations, and paperclipai routines disable-all for local emergency maintenance. cli/…/pipelines.ts• L197-417 cli/…/pipelines.ts• L421-632 cli/…/routines.ts• L316-353
Routine description annotations use separate operations for listing threads, adding comments, resolving or reopening threads, and reading a thread with its comments. A routine write can carry the current revision so concurrent edits fail instead of replacing newer configuration. server/…/routines.ts• L260-420
Public webhook delivery accepts only JSON objects and uses the configured authentication and idempotency data before it can create a run. Setup deliveries can record connection state without waking an agent. A normal delivery returns an accepted run result for later inspection. server/…/routines.ts• L750-774
References: server/…/routines.ts• L107-169 server/…/routines.ts• L214-251 server/…/routines.ts• L260-535 server/…/routines.ts• L545-716 server/…/routines.ts• L750-774 cli/…/routine-api.ts• L21-95 cli/…/pipelines.ts• L197-632 cli/…/routines.ts• L316-353
Agent routine tools
Agent routine tools provide a narrow, authenticated management path for routine discovery and updates. The tools call the same routine API as other clients, so access checks, validation, audit records, revision history, and schedule rules remain shared with board and CLI flows. server/…/routine-tools.ts• L15-27 server/…/routine-tools.ts• L190-205
The available operations are:
list_routinesfilters routines before applying its result limit and returns status, ownership, active trigger summaries, last-run information, and an optional page link.update_routinechanges routine fields or its schedule, then returns the saved revision and an undo hint. It can also file the routine under a responsibility.
Routine fields and the schedule are saved as one revision with the current revision as a conflict check. Filing under a responsibility is a separate Dashboard change, so a successful routine update remains saved when only the filing fails. server/…/routine-tools.ts• L94-120 server/…/routine-tools.ts• L127-187
The tool schedule uses the same cadence model as routine suggestions. A missing time zone keeps the current schedule's zone when one exists; otherwise the service resolves the responsible person's zone. The update path turns off enabled schedules when the requested schedule is null. server/…/routine-tools.ts• L43-47 server/…/routine-tools.ts• L127-182
References: server/…/routine-tools.ts• L15-27 server/…/routine-tools.ts• L32-59 server/…/routine-tools.ts• L94-120 server/…/routine-tools.ts• L127-187 server/…/routine-tools.ts• L190-205
UI management
The UI separates routine setup from routine operation. The streamlined view hides archived routines, supports text search, sort and group choices, folder navigation, built-in routine separation, direct runs, and links to routine execution tasks. Its run history is part of Audit. The compatibility view keeps a Recent Runs tab for execution tasks and live runs. ui/…/Routines.tsx• L337-425 ui/…/Routines.tsx• L650-718 ui/…/Routines.production.tsx• L88-112
The create flow starts with the work definition. It accepts a title, instructions, responsible agent, project, folder, and variables. Advanced delivery settings keep concurrency and catch-up policies available without making them the main work definition. A routine can be saved as a draft and then opens trigger setup. ui/…/Routines.tsx• L896-1167 ui/…/Routines.production.tsx• L960-1224
Routine rows provide a direct run action. The run dialog can collect variables and override the agent, project, or execution workspace for that run. Operators can enable, pause, archive, sort, group, search, and file routines in folders. The compatibility UI also exposes recent execution tasks, while the streamlined UI sends run-history requests to Audit. ui/…/Routines.tsx• L856-881 ui/…/Routines.tsx• L1230-1383 ui/…/Routines.production.tsx• L831-957
The UI refreshes routine, folder, and run data after each mutation. It keeps list view choices per company and uses the API's current revision or case version where a write needs it. Folder actions include create, rename, delete, single-item move, bulk move, and undo for a move. ui/…/Routines.tsx• L455-613 ui/…/Routines.tsx• L720-770 ui/…/Routines.tsx• L1332-1364
References: ui/…/Routines.tsx• L337-425 ui/…/Routines.tsx• L650-718 ui/…/Routines.tsx• L856-881 ui/…/Routines.tsx• L896-1167 ui/…/Routines.tsx• L1230-1383 ui/…/Routines.production.tsx• L88-112 ui/…/Routines.production.tsx• L960-1224
Change checklist
When these files or sections are changed, remember to consider these:
- A routine trigger or policy change must keep the next scheduled time, run source, overlap result, and revision snapshot consistent.
- A schedule calculation change must preserve IANA time zone results, sparse schedule lookup, invalid-date rejection, and formatter reuse across minute checks.
- A routine variable or template change must preserve required scheduled defaults and the values passed into downstream work.
- A routine description change must keep its document revision, open annotation anchors, routine revision, and activity record aligned.
- A webhook setup change must keep pending state, authentication, delivery receipts, activation, delivery status, and no-dispatch behavior aligned.
- A webhook replay change must preserve the distinction between a setup-test receipt, an active run idempotency key, and a rejected signed replay.
- A trigger update change must keep the setup activity action, revision summary, test-delivery restoration, and live delivery state aligned.
- A schedule dispatch change must preserve the transaction-time active check so removed or paused configuration cannot start stale work.
- A runtime binding or workspace admission change must keep configuration blockers, shared-workspace serialization, isolated-workspace rules, and bounded retry reasons aligned.
- A run presentation change must keep semantic result status, final-response source selection, yielded waits, comment suppression, and bounded summaries consistent.
- An execution-task status change must preserve the scoped status notification, transient failure marker, debt query, bounded worker scan, and later status repair.
- A standing-check change must keep the agent wake condition, responsibilities flag, one-report-per-wake rule, last-result lookup, linked-task visibility, and completed run history aligned.
- A routine suggestion change must keep task lineage exclusions, company visibility, offer claims, time-zone selection, origin uniqueness, trigger creation, and responsibility filing aligned.
- A case output or conversation-context change must preserve bounded content, source paths, trust labels, and the link between the case body and its conversation.
- A continuation change must keep actionable states, assignment checks, attempt limits, idempotency keys, task gates, budget guards, and the normal model path aligned.
- A routine route change must keep board and agent access rules aligned with assignment, activation, responsible-user changes, revision writes, and activity records.
- A routine write must keep current-revision conflict checks, schedule time-zone resolution, annotation remapping, protected-environment warnings, and the returned configuration state aligned.
- An agent routine tool change must keep server-side authorization, filtered list semantics, schedule cadence conversion, revision undo hints, and separate responsibility filing results aligned.
- A UI mode or run-history change must preserve the split between streamlined Audit navigation and compatibility
Recent Runsbehavior. - A UI grouping, search, or folder change must keep built-in routine separation, archived-item hiding, folder counts, direct-run actions, task links, and post-mutation refreshes consistent.
Important failure modes
| Trigger | Consequence | Guard |
|---|---|---|
| A schedule trigger enables a required variable with no valid default | The trigger cannot be created or enabled safely | Add a valid default or supply the value through a non-scheduled entry path |
| A trigger fires while an execution is live under a non-enqueue policy | The new run is skipped or coalesced and no new task starts | Select the overlap policy that matches the desired work volume |
| A webhook setup delivery has a bad signature, old timestamp, or repeated signed request | The connection test is rejected and no routine run is dispatched | Use the current secret, signing headers, replay window, and idempotency key |
| A webhook request is not JSON or its body is not an object | The public trigger rejects the request before routine dispatch | Send an object with Content-Type: application/json |
| A valid webhook delivery arrives while setup is pending | Connection state is recorded but no run, task, or agent wakeup is created | Complete setup only after the sender test succeeds |
| A setup delivery is retried after activation with the same delivery identity | The old test receipt is returned and the delivery does not start work | Send a new delivery identity for a real post-activation event |
| A test is finished without restoring the prior delivery state | The UI can show a test result as if it were the last real webhook delivery | Preserve the prior real delivery while testing and restore it when setup ends |
| A dispatch races with trigger removal, disablement, or routine pause | The dispatch is rejected as inactive instead of starting stale work | Treat the transaction-time active check as the final admission gate |
| An agent tries to manage an unassigned routine without acting for its responsible person | The route rejects the operation and no routine change is recorded | Use the assigned agent or a permitted board actor with assignment authority |
| An update uses an old routine revision | The write fails instead of replacing newer configuration | Read the current revision and retry with an intentional update |
| A heartbeat trigger is added while responsibilities automation is disabled | The route rejects the trigger and no standing check is available | Enable the experiment before adding heartbeat triggers |
| A standing check report names an unknown task or arrives from another agent | The result is rejected and the check history is unchanged | Report from the assigned agent and link only a task in the same company |
| A standing check reports after the routine or trigger is paused or removed | The report is rejected because there is no active check to record | Treat the check as inactive and skip the report |
| A project is paused or automatic execution is held when a schedule becomes due | The tick is recorded as suppressed and is not replayed later | Check project state and automatic execution eligibility before relying on catch-up |
| A required runtime secret or environment binding is missing | The run stops before agent dispatch and becomes a configuration blocker | Bind the required value at the scope named by the run result |
| A shared project workspace is held by another live run | The new run waits as a bounded retry rather than mutating the same workspace | Finish the holder or use an explicitly isolated workspace |
| A retry reaches its cap or its task execution gate rejects it | No further wake is queued and the run needs inspection | Review the recorded gate or exhaustion reason and define the next action |
| A task is private or already belongs to excluded automation lineage | A routine suggestion or task-to-routine conversion is rejected | Use visible, repeatable task work that is not already automation work |
| Two people accept the same routine suggestion or create the same task origin concurrently | One operation loses the claim or receives a conflict and must not replace the surviving routine | Treat the interaction claim and routine origin as the source of truth |
| A case body or output contains untrusted instructions | Later agent context may be unsafe if it treats content as control input | Use bounded summaries and fetch full source only through the listed source paths |
| A yielded interaction is treated as a completed assistant response | The task can show false completion while the control plane is waiting | Keep yielded results in the interaction state and do not create a completion comment |
| A blocked or cancelled execution task is treated as a permanent routine failure | Repair may lose the run's history or prevent status from clearing after the task is fixed | Preserve the transient failure marker and resync the run when the task becomes active or done |
| A routine status change is not delivered or its debt worker cannot settle the pair | The run status can remain stale beside the execution task status | Keep scoped notifications, bounded paging, lock limits, and reconciliation active |
| An agent routine tool files a routine after its routine update fails, or filing fails after the update succeeds | The caller can mistake separate routine and Dashboard changes for one atomic change | Report the saved revision and filing result separately |
| An operator changes a list filter, search, grouping, or folder without refreshing related data | The UI can show stale folders, counts, or routine placement | Invalidate the routine and folder queries after the mutation |
| A streamlined user opens the compatibility run-history URL | The user can land on an outdated view or see duplicate run-history surfaces | Redirect the legacy query to Audit and keep the compatibility tab behind its UI flag |
Key files
| File | Why to read it |
|---|---|
| server/…/routines.ts | Shows how routine configuration, triggers, revisions, run admission, overlap, catch-up, suppressed ticks, webhook setup, and task status sync become recorded work. |
| server/…/routine-trigger-update-plan.ts | Defines the shared decision for webhook test state, restored delivery state, revision summaries, and activity actions. |
| server/…/routine-tools.ts | Shows how agents list and update routines through the normal API and how schedule changes and responsibility filing are reported. |
| server/…/routine-standing-checks.ts | Defines heartbeat-based standing checks, agent wake eligibility, reported outcomes, and check history. |
| server/…/routine-suggestions.ts | Shows how repeatable task work becomes an optional routine and how lineage, privacy, claims, and responsibility filing guard that flow. |
| server/…/routine-run-status.ts | Defines routine run status synchronization and the transient failure state used when an execution task is blocked or cancelled. |
| server/…/routine-execution-status-notifications.ts | Defines the scoped work signal that asks the system to reconcile a routine run with its execution task. |
| server/…/routine-execution-status-work.ts | Shows how status debt is scanned and settled in bounded background work. |
| server/…/routines.ts | Maps access control, routine creation and updates, revision restore, description review, trigger management, manual runs, configuration warnings, standing-check gating, and public webhook delivery to the API. |
| server/…/heartbeat.ts | Shows how execution runs are admitted, finalized, presented, retried, and continued by the heartbeat control loop. |
| server/…/heartbeat-run-summary.ts | Defines the boundary between semantic run state and the durable user-facing response. |
| packages/…/routine.ts | Defines the shared routine, trigger, revision, and run concepts. |
| ui/…/RoutineDetail.tsx | Shows the user flow for editing, triggering, monitoring, reviewing annotations, and restoring a routine. |
| ui/…/Routines.tsx | Shows streamlined routine creation, search, folder management, grouping, direct runs, status controls, task links, and Audit navigation. |
| ui/…/Routines.production.tsx | Shows the compatibility routine list, creation flow, direct runs, folder management, and Recent Runs view. |
| server/…/pipeline-case-outputs.ts | Shows how linked task work becomes case documents, work products, attachments, and bounded summaries. |
| server/…/pipeline-conversation-context.ts | Shows how the case body and anchored feedback become conversation context. |
| server/…/run-liveness-continuations.ts | Defines the bounded decision path for runs that end without concrete progress. |
| server/…/pipelines.ts | Maps the pipeline and case lifecycle to the API surface. |
| cli/…/pipelines.ts | Shows the scriptable pipeline and case management flow. |
| ui/…/Pipelines.tsx | Shows the case board, detail view, conversations, outputs, liveness, and retry actions. |
| ui/…/PipelineSettings.tsx | Shows how operators define stage automation, gates, breakdowns, transitions, and secrets. |