sweep-labs

Find connect-labs artifacts (workflows, pipelines, synthetic opportunities, solicitations, funds, reviews, responses) that no current opp references. Workflows, pipelines, synthetic opps, and solicitations have first-class delete/disable atoms in the labs MCP — deletable in the execute pass, after the orchestrator's per-system human-confirmation gate (see agents/sweep.md § Human-confirmation gate). The remaining LabsRecord-backed types (funds + standalone reviews/responses) are report-only by design until per-type atoms exist upstream. Reviews and responses cascade-delete with their parent solicitation, so they generally don't need standalone cleanup.

Inputs

• Live-set file path from sweep-live-set.
• LABS_MCP_TOKEN and labs base URL (loaded via connect-labs MCP).
• Labs admin base URL: https://labs.connect.dimagi.com.
• mode — recommend (default) | execute. In recommend the skill is report-only: diff/score/render + return the recommended-action list; no mutations (no delete/disable). In execute it mutates only approvedIds. The human-confirmation gate between the two is the orchestrator's job — see agents/sweep.md § Human-confirmation gate.
• approvedIds (execute mode only) — the exact workflow/pipeline/synthetic/solicitation/update-link/quote ids the human approved in chat. Mutate nothing outside this set. A force: true delete of a solicitation with seeded responses is approved per solicitation, by name, never by a blanket approval.
• programIds (optional explicit argument) — labs program ids to audit for supply-chain records. Default: every program_id on a solicitation ACE's orgs own (Step 2), plus every phases.execution-management.products.supply_chain.program_id the live-set's runs carry.

Products

• ACE/_sweep/<timestamp>/labs-orphans.md — human-readable triage report.
• ACE/_sweep/<timestamp>/labs-orphans.yaml — machine-readable OrphanReport.
• For approved orphan workflows: workflow_delete calls.
• For approved orphan pipelines: pipeline_delete calls.
• For approved orphan synthetic opps: synthetic_disable calls.
• For approved orphan solicitations: delete_solicitation calls (cascade-deletes reviews + responses; gated on cascade emptiness — connect-labs PR #197).
• For approved supply-chain items: supply_chain_update_link_revoke and supply_chain_quote_void calls.
• An irreversible-records section in labs-orphans.md: the supply-chain records labs keeps no delete for, listed for audit (§ "Optional labs track" below).

Process

1. Read the live-set via drive_read_file.
2. List labs inventory using existing atoms:
   • workflow_list
   • pipeline_list
   • list_solicitations (each entry includes program_id, status and is_public). It returns every call ACE's organisations own, unlisted (is_public: false) ones included, plus every public call. No program_id is needed to see an unlisted call you own — that was a stale claim, refuted against the live schema in #3166. Pass organization_id per ACE org (pm_org, nm_org, llo_org) to scope the listing to ACE's own calls.
   • list_funds
   • list_reviews
   • list_responses
3. Diff against the live-set's labs buckets:
   • workflows → liveSet.identifiers.labsWorkflowIds
   • pipelines → liveSet.identifiers.labsPipelineIds
   • synthetic opps → liveSet.identifiers.labsSyntheticIds
   • solicitations/funds/reviews/responses → liveSet.identifiers.labsRecordIds
4. Score each orphan via scoreLabsItem(item, liveSet) from lib/sweep-fingerprint.ts. Pass each item's name AND its description (workflows, registries, cohorts) / notes (synthetic opps): a Phase 7 object carries its sweep handle (synthetic, ace <opp>/<run-id>) in whichever field no frame renders — a rendered object (programme report, opp report, synthetic opportunity) has it in description / notes, never its name (demo-data-setup § C7 "Sweep handles", ace#2733). sweepHandleField checks all three and the score names the field and the run it came from. The scorer defensively downgrades to low when an item references a Connect opportunity that's still in liveSet.identifiers.connectOpportunityIds — that's a sign the caller mis-flagged the item.
5. Build the OrphanReport with system: 'labs'. Partition into:
   • Actionable (deleted in the execute pass, after human approval): workflows, pipelines, synthetic opps, orphan solicitations. Solicitation deletion is gated on responses + reviews == 0 (cascade emptiness), not on lifecycle status — every ACE dogfood solicitation passes that gate cleanly because no real LLO ever engaged. See connect-labs PR #197.
   • Informational only: funds, standalone reviews/responses (reviews+responses that were already cascade-deleted with a parent solicitation don't appear here). Funds may need their own per-type atom in a future labs MCP release; reviews/responses generally cascade with their parent solicitation.
   • Supply-chain items (optional labs track): active update links, and quotes that are neither awarded nor voided, in a program that is orphaned or whose run has closed — see § Optional labs track.
6. Render + write labs-orphans.md and .yaml to the sweep folder, including the irreversible-records section.
7. Recommend — stop here in mode: recommend. Return the structured recommended-action list (orphan workflows/pipelines/synthetic to delete-or-disable; orphan solicitations to delete with their cascade preview — each with id, confidence) plus the informational-only items (funds, standalone reviews/responses) and the report Drive link, to the orchestrator. Perform no mutations. Do not try to prompt the human from this skill — a dispatched subagent can't reach them; the orchestrator runs the confirmation gate (see agents/sweep.md § Human-confirmation gate).

Execute phase (mode: execute only)

Runs only when the orchestrator re-dispatches with mode: execute + approvedIds (the ids the human approved in chat). Mutate only those ids:

• Orphan workflows → workflow_delete({ workflow_id })

• Orphan pipelines → pipeline_delete({ pipeline_id })

• Orphan synthetic opps → synthetic_disable({ synthetic_opp_id })

• Orphan solicitations → delete_solicitation({ solicitation_id }) (program_id is "accepted for compatibility; not needed" in the live schema). The atom cascade-deletes the solicitation's reviews and responses in one call and returns the count of records deleted at each level — surface that in the result so the human sees the cascade impact. Pass force: true only for the rare legitimate destroy-with-engagement-data case (test fixtures that intentionally include responses) — and only if the human explicitly approved that override.

• Approved update links → supply_chain_update_link_revoke({ program_id, link_id, source: { ref: "ace:sweep:<timestamp>:update-link:<link_id>" } }). It stops working at once and cannot be re-enabled; what was recorded through it stays.

• Approved quotes → supply_chain_quote_void({ program_id, quote_id, reason: "ACE sweep <timestamp>: simulated run <run-id> closed", source: { ref: "ace:sweep:<timestamp>:quote:<quote_id>" } }). The quote stays readable and drops out of comparisons.

Funds, standalone reviews/responses and every irreversible supply-chain record are never mutated (report-only). Return the per-item result.

Optional labs track (--with labs_e2e runs)

A run with the optional steps on (spec docs/superpowers/specs/2026-10-18-optional-labs-e2e-steps.md) leaves three kinds of labs objects behind, each with a different cleanup:

1. Private solicitations with seeded responses. solicitation-seed-responses files fictional respondents (names end [Synthetic], the run's roster lists them). Because the call then HAS responses, delete_solicitation refuses it without force: true. Recommend it as a per-solicitation force delete, with the cascade preview, and say the responses are ACE-seeded fixtures. The human approves the override for that solicitation by name. A solicitation with any response NOT on the run's seeded roster is real engagement data: report it, never recommend force.
2. Removable supply-chain items. For each program in programIds: supply_chain_update_link_list({ program_id, include_inactive: false }) → recommend revoking every active link of an orphaned or closed run; supply_chain_quote_list({ program_id }) → recommend voiding every quote that is neither awarded nor already voided, on a tender no live run still has open. These are the only supply-chain removals labs offers.
3. Irreversible records — report, never act. Labs keeps no delete for tenders, quotes, awards, contracts, shipments, consignments, receipts, invoices, payments, stock counts or distributions (spec § F6), and a declared programme runner can only be changed by that runner's admins or support, never removed. For each program, list counts from the matching supply_chain_*_list reads and supply_chain_programme_runner_get, each with the run that wrote it. The isolation for these is the program itself: they are scoped to ACE's own program, named "simulated", and stay there for audit.

Known ACE-made items as of 2026-10-18, so the first sweep after them reads them right rather than flagging them as strays:

• Private call 1000003 on program 308 (owned by connect_orgs.pm_org), closed, with seeded responses 1000046–1000049: the PR4 live probe (#3176). Deletable: recommend force: true with the cascade preview.
• Program 308's runner declaration (connect_orgs.pm_org's labs org, 352): the live check of connect-labs #2605. Irreversible from ACE; report it. Nothing else was written to program 308's supply chain.

Failure modes

• Live-set path doesn't resolve: abort with "Run sweep-live-set first."
• Labs MCP returns 401: LABS_MCP_TOKEN is expired; recommend /ace:labs-token-mint to rotate, then retry.
• workflow_delete fails because the workflow has running tasks: report the item as "in-use — needs manual review"; don't retry, don't abort the batch.
• delete_solicitation refused (FAILED_PRECONDITION with response/review counts): the solicitation has real engagement data — a real LLO filed a response or a review was created. Surface the counts in the chat output; do NOT pass force: true automatically. The operator decides whether to override.

Why funds + standalone reviews/responses remain report-only

The labs MCP exposes per-type delete for solicitations because solicitations are the lifecycle unit for the audit-data ecosystem (reviews and responses are children that cascade-delete with their parent). The other LabsRecord types intentionally don't have per-type delete atoms because:

• Funds may have allocations + downstream allocation refs from responses. A future delete_fund atom would need similar lifecycle guards (and the labs team hasn't surfaced demand for one yet — funds tend to be reused across solicitations rather than created per-opp).
• Standalone reviews/responses (not cascade-deleted with a parent solicitation) are uncommon — if they exist as true orphans, they were probably created by a buggy producer and should be investigated rather than auto-deleted. Surface them for visibility.

If the labs team adds delete_fund upstream, this skill's coverage matrix entry flips automatically — the proxy is a pure forwarder so no ACE-side wiring is needed beyond updating Step 5/6/9 and agents/sweep.md.

Related skills

• sweep-live-set produces the live-set this skill diffs against.