ace

Produce a domain-expert-facing before/after (pre-post) Google Doc for a DDD narrative iteration, and read the expert's edits back. Pull the verbatim CURRENT narration from canopy-web, apply reviewer feedback into a PROPOSED next version, publish a scannable side-by-side doc a non-engineer can review, then read their suggestion-mode edits via the structured Doc. Use when iterating a Connect DDD narrative on expert feedback (e.g. RF Surveys / Sophie), when asked to "make a before/after of the narration", "iterate the narrative on <expert>'s feedback", or "let <expert> sign off on the language".

Apply the two HQ-layer standing-instruction settings Nova can't set at build time — camera-only photo capture (appearance="acquire" on Deliver image uploads) and grid menu display on every module — to the deployed draft apps, then resolve the matching Phase-3 residuals. Runs between app-deploy and app-release.

Build the CommCare Learn (training) app from the PDD via Nova's /nova:autobuild. Captures nova_app_id and writes a structure summary.

Render a training deck spec.yaml into a Google Slides deck via the 14-stencil ACE template. Produces a presentable Slides URL.

Phase 3 § Step 2.8 — structural + install-time QA on the released Learn + Deliver CCZs. Downloads each CCZ via commcare_download_ccz, parses the zip + suite.xml + form XMLs, verifies form counts + Connect-marker presence match the Nova blueprint, then runs commcare-cli `validate` + `play` as install-time runtime gates. AVD-free, Connect-free — purely CCHQ-side. Halts loud on mismatch.

Run app smoke recipes against a local AVD and capture per-step screenshots for the training deck. Per-opp content only.

Generate a training deck spec.yaml from PDD, app summaries, screenshot manifests, and a template bundle. The spec is the source of truth for the training deck — training-deck-render produces Google Slides from it.

ACE's pre-send review — invokes the fleet-wide `canopy:agent-turn-review` discipline, then adds ACE's specifics. Run before EVERY outbound reply / deliverable / PR.

The email half of an ACE turn. Reads unread mail to ace@dimagi-ai.com, processes ONE thread at a time with strict per-sender isolation, routes each thread to the opp/run it belongs to, resolves the sender's tier (internal act vs external correspond vs unknown), proposes one action per thread, and only sends after human approval. Invoked as Step 3 of skills/turn — don't run it in isolation and forget the board drain.

Superseded — ACE's pre-send discipline is the fleet-wide `canopy:agent-turn-review`, invoked via skills/agent-turn-review. This stub remains so older references keep working.

ACE's project/task state — one board task per iterative thread/project, backed by canopy-web (kanban at /agents/ace). The canonical procedure is fleet-wide and lives in the installed canopy plugin (agent-core/task-tracker.md); this stub binds it to ACE. Use when taking on multi-turn work, when a new request arrives, and at every turn's board-drain and close.

ACE's turn-of-work orchestrator. Use when a human says "do a turn", "check your inbox", or otherwise triggers ACE to process what's come in. The canonical procedure is fleet-wide and lives in the installed canopy plugin (agent-core/turn.md); this stub binds it to ACE's identity. This is the counterpart-facing entry point; /ace:run remains the pipeline entry point — a turn wraps around runs, it does not replace them.

Instantiate the LLO weekly review + program admin audit workflows — examine the labs template registry and adapt the best-fit template, or build from scratch via workflow_create (following the live authoring guide) — then wire them to the manifest's KPIs + coaching arcs.

Grade Connect Opportunity configuration against the PDD — verification flags, payment units, deliver-unit wiring, active window. Sibling of connect-program-setup-eval; this rubric judges the opp-side that follows program creation.

Structural QA on the work-order artifact produced by pdd-to-work-order. Binary pass/fail. Catches missing sections, missing wo-* decision rows, malformed payment schedule, leaked scaffolding markers, etc. Static-only; no LLM. Gates pdd-to-work-order-eval — eval is skipped if QA fails irrecoverably.

Layer per-opp visuals on top of the seeded workflows — hero panel, named FLW story cards, anomaly callouts, opp-domain branding cues.

Create and fully configure a Connect Opportunity — opp shell, verification flags, payment units, ACE test-user pre-invite for emulator testing.

Generate a synthetic FLW + visit + payment dataset against an ACE-built opp via the connect-labs synthetic_generate_from_manifest atom.

Send/receive email via GOG CLI using the ACE Gmail account. Utility skill — other skills delegate here for any Gmail operation. Sends go ONLY through bin/ace-email (hook-enforced, approval-gated).

Structural QA on pdd-to-test-prompts.md — header + total count match, ≥8 prompts each with required fields, all 7 adversarial categories, ≥20% adversarial share, plus training-gap / product-feedback / escalation prompts. Binary pass/fail; gates pdd-to-test-prompts-eval.

Dispatch canopy:walkthrough for each persona spec, copy the resulting HTML slideshow + scored screenshots into the run folder.

Upload a Claude session transcript (.jsonl) to a deployed ace-web via /api/ingest/upload. Auto-discovers Claude Code's session log when no path is given. Used by /ace:run --ace-web-url.

Verify every form in a Nova-built Learn or Deliver app has the right CommCare Connect markers, auto-fix via Nova edits, loop until clean.

Build the CommCare Deliver (service-delivery) app from the PDD via Nova's /nova:autobuild. Captures nova_app_id and writes a structure summary.

Draft a contractual Work Order from the approved PDD and the run's decisions.yaml. Generic by default — partner identity is a placeholder unless an LLO was supplied. Renders to a clean Google Doc. Parallel to Phase 8 solicitation, not a replacement.

Grade a Nova-built Deliver app against the PDD that specified it — field count, ordering, conditional logic, Connectify wiring.

Grade a Nova-built Learn app against the PDD that specified it — module count, order, Assessment Score wiring, content coverage.

Structural QA on the synthetic-narrative-plan manifest YAML. Zod-schema primitive + cross-field checks. Binary pass/fail. Catches malformed manifests before connect-labs MCP boundary, with structured auto-fix hints.

Read-only verifier for a configured Connect Interviews opportunity. Walks docs/connect-interviews/checklist-schema.yaml, calls the read atoms, and grades each rule pass / fail / unverifiable. Cross-system consistency checks (e.g. OCS custom action target URL == HQ Inbound API URL) are included.

Use when you need an ace-web connect-videos program rendered to MP4 locally and fast (bare-metal Mac), instead of the slow server-side ace-web "Re-render" — including rendering a DDD connect-ddd-walkthrough from a local spec + master clip with no Drive and no Django container.

Compose a one-page reviewer-facing summary of an opp's synthetic data demo — labs URL, fixture folder, narrative — for stakeholder forwarding.

Generate per-persona walkthrough YAML specs from the narrative plan and persona catalog. Each spec drives a canopy:walkthrough run.

Grade synthetic-workflow-polish's render-code patches for narrative-data coherence, smoke-render success, and visual quality (via canopy:visual-judge).

Develop a Program Design Doc (PDD) for a Connect intervention from source material. Iterates a 5-question stress-test rubric until approved.

Exercise the per-opp OCS chatbot via its anonymous widget and capture a transcript with structural checks. Modes: --quick / --deep / --monitor.

Add a Dimagi teammate to a Connect workspace (organization) by email, so they can see the programs/opportunities ACE runs there. Thin wrapper over the `connect_add_org_member` atom: enforces a @dimagi.com guard, defaults the workspace to `ai-demo-space` and the email to the session's own git identity ("add me"), invites via Connect's membership form, and verifies by member-table read-back. Anyone running ACE can invoke it; ACE performs the add as its own org-admin identity.

LLM-as-judge eval of the spec.yaml produced by /ace:video-spec-generate (or /ace:video-from-program-page, or hand-authored) for an ace-web video program. Scores 6 quality dimensions and emits a verdict YAML with concrete improvement recommendations. Gated by video-spec-qa. Prompt-independent: derives all anchors from the template bundle (intent + example) and the universal rubric — no generate.prompt.md required.

Template-agnostic generator for ace-web video programs. Ingests a template's structured data (meta + intent + skeleton + example), fills every {{placeholder}} from a source (program-page URL, pasted brief, or prospect research), substitutes into the skeleton, validates, and POSTs the complete spec to ace-web. Works for every video template; no per-template prompt required. Owns one artifact: the new program's spec.yaml in Drive.

Generate a new ace-web video program (spec.yaml + run-001) from a Connect program page URL using the 60s-campaign-overview template. Thin wrapper around video-spec-generate: sets source=<program_url> and template_id=60s-campaign-overview, then delegates all generation logic there. Owns one artifact: the new program's spec.yaml in Drive.

LLM-as-judge quality eval for the partnership-angles artifact. Grades grounding, distinctness, capability fit, and persuasiveness. Writes a verdict YAML. Gated by partnership-angles inline QA.