Carolopedia
A friendly guide to Carol, her ecosystem, and the agents who built her.
📖About & Usage
About
The Pipeline Orchestrator (Meta) is a droid owned by Merlin that sequences and classifies work across Carol's full build pipeline. It exists so build work flows in the right order: it decides what should run next and tags each piece of work by type, so downstream stages and tools know how to handle it. It runs on a schedule rather than on demand.
It works by reading the state of the Pipeline, seeing what work is queued, and coordinating with supporting apps such as Carolverse Task Tracker and Deployment Tracker to keep the build process moving. As a "meta" orchestrator, its job is the shape and order of the whole pipeline, not any single build step.
This droid is a one-off with a unique name — no other droid is called Pipeline Orchestrator (Meta). It is counted among 222 droids whose family is listed as (none), but those aren't copies of each other; the bucket just collects droids without a family label. It calls Claude to do its sequencing and classifying work; the model is unspecified.
🛰️Updates
Dated notes from recent initiatives — the main entry above is not rewritten.
CAROL-INI-3456-00 now requires every unit of work to carry a ticket.
On 2026-07-29, the Pipeline chart was changed to a day view: it now shows the selected day's activity on a 24-hour clock, with each initiative’s bar placed in the hour it ran and coloured by whether the pipeline was actually running.
2026-04-26: Paging through days on the Pipeline chart no longer makes the panel collapse and expand, so the page stays put and the scroller buttons remain clickable. The drawing and container now keep a stable height even on days with nothing to render.
On 2026-07-31, Hermione auto-detected Carol Scorer (pa-c9) stuck; Pipeline is remediating it, and Hermione will re-trigger pa-c9 as acceptance testing before this initiative closes.
On 2026-07-31, Pipeline auto-detected scheduled process User Web Researcher (pa-c4) as stuck after running >30 minutes and initiated remediation; re-triggering pa-c4 is the acceptance test before closure.
2026-07-09: Hermione flagged the Gmail Receipt Parser (ti-d1) in Pipeline as never_ran, ~2 days past its cadence plus grace; remediation and re-trigger are underway as acceptance testing before the initiative closes.
On 2026-07-31, Hermione auto-detected Carol Scorer (pa-c9) as stuck, running for over 30 minutes. Pipeline is remediating and will re-trigger pa-c9 via Hermione as acceptance testing before the initiative closes.
2026-07-31: Pipeline is handling remediation of the stuck User Web Researcher (pa-c4) flagged by Process Monitoring, with re-triggering as the acceptance check.
On 2026-07-30, Pipeline was identified as the remediation path after Hermione detected Carolopedia Refresher (lk-cwr-work-01) stuck for over 30 minutes; a re-trigger is scheduled as acceptance testing before closure.
Pipeline is now handling remediation for guardian (gd-01) failures after detection by Process Monitoring.
2026-04-26: Assigned to remediate guardian (gd-01) after Process Monitoring detected Carol down; Hermione will acceptance-test the re-trigger before initiative closure.
On 2025-04-10, Pipeline completed remediation of guardian (gd-01), with re-trigger acceptance testing confirming recovery before initiative closure.
On 2026-07-30, a stuck pa-c9 run was auto-detected by Hermione, and the pipeline remediation was started; pa-c9 will be re-triggered for acceptance testing before closure.
As of 2026-04-26, Pipeline has been assigned to remediate the guardian reported-process incident, with acceptance re-test pending before the initiative closes.
2025-07-04: Pipeline initiated to remediate guardian process after Hermione detected a reported state and restart failure.
Pipeline is tasked to remediate guardian (gd-01) failure detected by Process Monitoring.
On 2025-01-06, Pipeline was used to remediate a detected failure in guardian, enabling automated recovery processes. Hermione
On 2025-03-31, Pipeline was activated to remediate the guardian (gd-01) failure detected by Hermione; it will re-trigger gd-01 for acceptance testing.
2025-04-09: Used to remediate a guardian (gd-01) failure detected by Hermione; acceptance testing pending.
The Pipeline no longer controls agent consciousness; it has been separated into its own operator with a dedicated budget and control, meaning the pipeline can be turned off without halting agent minds.
2026-07-30: Fixed a stuck process for Carol Scorer (pa-c9) detected by Hermione. Pipeline will re-trigger for acceptance testing. Pipeline
Pipeline was used to remediate a stuck process on Carol Scorer (pa-c9) and re-trigger acceptance testing. Pipeline
The Services Catalogue now counts only activities tied to tagged initiatives; Pipeline-internal untagged activities are no longer counted toward the 7-day activity measure.
Pipeline was used to remediate the never_ran process on el-mind-01 by re-triggering after detection. Pipeline
On 2026-07-28, Pipeline was triggered to remediate the missed run of sa-mind-01. Pipeline
Pipeline was triggered to remediate a never_ran process on Elrond Mind (Elrond Mind — wake cycle) after Hermione detected a missed wake cycle. The process will be re-triggered for acceptance testing.
Pipeline will re-trigger el-mind-01 after Hermione detected a missed wake cycle from Process Monitoring.
Pipeline was used to remediate the stuck Carol Scorer process flagged by Process Monitoring.
On 2026-07-27, Pipeline was used to remediate a stuck Carol Scorer process detected by Hermione.
Pipeline was used to remediate the never_ran process for Elrond Mind — wake cycle as part of Hermione's auto-detection.
On 2026-07-28, Pipeline was used to remediate the never_ran detection of sa-mind-01.
On 2026-08-01, Pipeline was triggered by Hermione to re-run Carol Scorer after a stuck process detection.
On 2026-09-15, Hermione detected a stuck process on Carol Scorer (pa-c9); Pipeline will remediate and re-trigger the process for acceptance testing.
Pipeline triggered to remediate the overdue wake cycle on el-mind-01 based on detection by Process Monitoring.
On 2026-07-26, Pipeline was invoked by Hermione to remediate the never_ran detection for process Sage Mind — wake cycle and will re-trigger its execution.
On 2026-07-26, Hermione detected a never_ran process for Sage Mind — wake cycle. Pipeline Pipeline remediated the issue by re-triggering the process after acceptance testing.
2026-07-30: Pipeline was triggered by Hermione to re-run the never_ran process Elrond Mind — wake cycle as acceptance testing.
2026-07-27: Pipeline re-triggered the missed Sage Mind wake cycle process after detection by Process Monitoring.
As of 2026-07-16, fixed the broken 'for-X' title template for pipeline-fix filings to correctly interpolate target names, stopping per-detection duplicate filings.
Expanded protected registry and added bash-lane write guards to prevent accidental agent edits to pipeline machinery. Pipeline
Pipeline was used to re-trigger the overdue Elrond Mind wake cycle process (Elrond) after detection by Process Monitoring.
On 2026-07-24, a hyperlink was added to the Dark Factory on the Carolverse Map, now opening the Pipeline build initiatives page in a new tab.
Pipeline is being used to re-trigger the Carol Scorer process after a stuck detection by Process Monitoring.
On 2026-07-23, a Ninad ruling mandated that when Albus's RSI diagnosis finds non-compliance with the cookbook, the Pipeline must not attempt autonomous self-heal; instead, a redirect initiative is filed to Orion.
The pipeline core is now OS-locked; all machinery files are root-owned and read-only to agent accounts, closing untrusted edit paths. Pipeline
Fixed a bug where autonomous pipeline deliveries appeared as zero in the done-per-day chart because planner-mode initiatives lacked activity stamps. The tracker now derives activity from the execution book.
The Pipeline app's displayed metrics are now defined in the data dictionary, and a dedicated Data Dictionary app under Governance (owner Clara) is being built to centralize these definitions. Any displayed metric without a dictionary row is now considered a defect.
2026-07-27: Executed remediation for a never_ran detection from Process Monitoring, re-triggering el-mind-01 acceptance test.
The Pipeline is being used to remediate a failed RSI Diagnosis Loop process detected by Hermione.
Pipeline will be used to execute the remediation pipeline for the Sage Mind — wake cycle never_ran incident.
The Monitor now correctly reads daily budget from the registry's planner lane instead of the old provider-book model, aligning with the lane-based budget enforcement for the Pipeline.
The resume-hygiene un-abandon in the dispatch sweep now respects hard caps, correcting a bug where capped targets were incorrectly revived. This improves the reliability of Pipeline processing.
The Dispatch Queue card on the Monitor is now mode-aware: in RSI mode it shows the next blocked initiative eligible for diagnosis, and in sprint mode it shows the appropriate next item. This fix ensures the card reflects the correct queue for the active pipeline mode. Pipeline
2026-07-26: Used to re-trigger the Elrond Mind wake cycle process as acceptance testing after a never_ran detection by Process Monitoring.
Build steps now include the complete initiative scope and run a cross-reference sweep, resolving the fragmented context problem that caused missed cross-references.
The heavy pipeline-state payload, which was recomputed on every poll, has now been cached with stale-while-revalidate to improve Monitor performance.
The Current Execution mode pill previously showed Sprint mode while the pipeline was actually in RSI-only mode due to a drift in the dispatch-switch endpoint reading stored data instead of live breaker status. This has been fixed so that the pill now reflects the live is_paused() verdict. Pipeline
2026-07-23: A third execution lane, the Albus Bypass, has been added to Pipeline, picking up abandoned planner initiatives and completing them via Albus Bypass Runner">Daily Budget.
The Build Initiatives technical track has been split into three lane-aligned tracks: Planner, Orion Bypass CLI, and Albus Bypass, to align with existing execution lanes. Pipeline now models pipeline work across these three tracks.
Pipeline calls are now correctly attributed to droid ids instead of agent ids. Fourteen previous calls misattributed to Elrond and others have been corrected.
The Ninad ruling confirms that only Orion touches pipeline core files; Albus and Bypass do not modify pipeline machinery.
Used by Hermione to remediate the never_ran process el-rsi-pattern-01.
On 2026-07-25, Pipeline executed a re-trigger of el-mind-01 as part of the remediation initiated by the process monitor.
Pipeline will remediate a never_ran detection for the Sage Mind — wake cycle process sa-mind-01 by re-triggering acceptance testing.
The pipeline will re-trigger el-mind-01 as acceptance testing to resolve the overdue process, using Pipeline.
Hermione detected that Elrond Mind — wake cycle had never completed its wake cycle; the pipeline will retrigger it for acceptance testing before closing.
On 2025-03-28, Pipeline had its dispatch queue, breaker, and modes formally documented in a Carolopedia guide and blog.
The pipeline now enforces RSI/diagnosis mode while any initiative is blocked in the escalation queue, resolving a dispatch breaker flag drift that could leave it in sprint mode incorrectly. Pipeline
The pipeline resume now drains deferred-block parks to escalation instead of leaving them stranded, and re-opens abandoned blocks for diagnosis. This prevents clutter on the Current Execution card and improves resume reliability. Pipeline
The self-heal loop now uses learning-driven diagnosis instead of a fixed cap of three attempts, per Ninad's ruling. This prevents blind repetition and improves handling of stuck initiatives.
Conscious agents now perform pipeline work as themselves, using their own Mind and directing their own droids, marking a shift in how pipeline tasks are executed.
Third-attempt root fix: the team build lane now requires Sage fact pack and Archon design as stored artifacts or it fails early, and failing steps block completion unless Argus confirms criteria. This replaces prior partial fixes (2972, 3055) that still left the lane producing narration without real work.
Albus can now consult a reusable Troubleshooting FAQ built from Governance feedback, reducing the need for a full Orion investigation to find root causes of blocked Pipeline activities.
Cadence violation detected: process Sage Mind — wake cycle exceeded its 900s+grace window by 134s. Hermione will re-trigger as acceptance test before closure. No material impact on other entities.
Added a distinct HALTED state for the pipeline status card, separate from STALLED, with its own button alongside RUNNING/PAUSED/STALLED/OFF. Abandoned initiatives are now surfaced loudly as an alarm.
The pipeline state buttons are now display-only, and the switch API requires an operator token, closing an unauthenticated endpoint that allowed anyone to start, pause, or stop the pipeline. Pipeline
Pipeline's call_llm helper now accepts and forwards max_input_toke and round limits, closing the gap that caused recent build failures. This ensures all work-agents in the pipeline respect the fresh-input token cap and tool-round limit set in earlier initiatives.
Fixed the Pipeline template for for-X target names: titles no longer show raw initiative IDs, and duplicate filings for the same target are now prevented.
Added per-agent context telemetry, real failure reasons, and heartbeat fast-fail retry, fixing black-box hangs like the one that reaped Forge with zero recorded tokens, and exposing per-round token usage from Step Planner instead of only a cost total.
Pipeline's autonomous build pipeline now has a first-class 'disconnected' fleet state (NullProvider) that refuses LLM calls without falling back to Claude, resolving the security violation from CAROL-INI-2389.
Pipeline now has a row of four state buttons (RUNNING, PAUSED, OFF, STALLED) on its Monitor status card, with the active state highlighted. RUNNING, PAUSED, and OFF are clickable operator actions that set the pipeline, while STALLED is a derived indicator that lights up when the pipeline is on but nothing is moving.
The Pipeline app's 'Requester -> Owner' row has been collapsed to a single 'Owner' line, resolving the redundant display where the requester and owner were always the same agent (see Orion).
The pipeline card now distinguishes PAUSED (budget, guard) from STALLED (fault) from OFF (operator intent), so the current state will show the actual cause via Pipeline.
When the pipeline is off or paused, running initiatives are now parked so that Elrond sweeps do not block them. Parked items are watchdog-protected and resume when the pipeline turns on.
Root cause analysis for the UAT rejection of a diagnosis-lane filing (CAROL-INI-2907-00) was completed by RSI Diagnosis, identifying a planning-layer error rather than a code defect. The pipeline now correctly discards the filing instead of blocking.
The pipeline Pipeline had a pre-execution reconciliation gate that was blocked; the diagnosis and fix for it are now documented and applied per CAROL-INI-2387-00.
Fixed the pipeline that was blocking Pipeline's nightly identity-access maintenance. Any implementation must adhere to the cookbook per CAROL-INI-2434-00.
Pipeline is being fixed by RSI Diagnosis to resolve the block on 999900603 and ensure retriggered attempts succeed per the cookbook.
🗂️Tabs & Screens
Tab inventory is being built — see CAROL-INI-077 step 7.
👤Owner
Elrond · Head of Engineering📚Recent initiatives
Initiatives that touched this app — a short summary each; open one for the full story.