Aller au contenu principal

Open Questions — Agents, Skills, Tasks specs

Round-9 status update (2026-05-25) — design set locked. Operator answered all remaining questions. Most defaulted to the ★ recommendation; explicit round-9 overrides:

  • N5 override — BOTH export AND import for Agents ship in v1 (was ★a = export only). New phase 6a (T64a-T64h) in agents/tasks.md for the per-Agent envelope + controller + UI + e2e. Distinct from the bulk account-transfer flow in ADR-008.
  • N6 override — Keep all 5 AgentBudget.intervalUnit values in v1 + implement multi-interval aggregator. Reversed round-6 narrowing in agents/plan.md §3.1; new task T34a covers the getCurrentPeriodStart / getNextPeriodStart helpers handling hour/day/week/month/unlimited with anchored rolling periods.
  • P1 — 6 starters seeded in ever-works/agents repo at launch: CEO, CTO, Researcher, PR-Reviewer, Editor, Designer.
  • P4 explicit no-modal — No announcement modal, no migration story, no existing-user UX work. Operator quote: "NO NEED THIS! WE DO NOT HAVE ANY REAL USERS YET, WE DON'T NEED SUCH THINGS!!!" Confirms K3. UX-DESIGN §11.3 updated.
  • All other items are now [Answered: ★ default per recommendation] — bulk-marked by script. Spec set is design-locked; next phase is implementation per the 12-18 PR shipping plan in architecture/implementation-reuse-map.md §14.

Round-8 status update (2026-05-25): Operator answered the bulk of A-L sections inline. The [Answered: X] tag now appears on each resolved question with the chosen option. Where the answer overrode my ★ recommendation OR triggered substantive spec changes, the change is noted in the relevant feature spec / plan / tasks file with a back-reference to this question. Major substantive impacts from round-8 answers:

  • F4-b (reversed my F4-a): Tasks created on an Idea do NOT follow to a Work when the Idea is accepted. Idea-level tasks (validity checks, marketing efforts, etc.) stay on the Idea.
  • F5 (operator override): Recurring tasks MUST ship in v1 (was Out of Scope). New Trigger.dev cron task; full plan in task-tracking/plan.md.
  • F3 clarification: Tasks ≠ Ideas. Ideas are spec/PM "Epic"-ish abstractions (auto-generated, droppable, never bugs); Tasks are implementation-level work units. Added to task-tracking/spec.md §1 and ADR-009.
  • H3 (operator override): Agents have THREE avatar modes in v1 — initials (default, auto-derived), icon picker, image upload (when tenant storage enabled). Updated UX-DESIGN + Agents spec + plan.
  • K1 (yes, draft): New doc architecture/agent-yml-manifest-schema.md ships the Zod schema for agent.yml.
  • K2 (yes, but trust): UX-DESIGN gets light additional mockups for the most novel surfaces (Tasks Kanban detail, Agent dashboard live state).
  • K3 (no migration): we don't ship a migration story for existing users; as long as nothing breaks for them, we're fine.

Unanswered yet (∼50 items remain open across sections M / N / O / P / Q / R3 / S): these default to their ★ recommendations unless the operator overrides. Marked [Open] in the headings below.

Round-6 status update (2026-05-25): Operator answered several open questions inline during PR review. Resolved items:

  • A2 (Skill catalog placement): separate repo ever-works/skills via plugin → ADR-007 superseded by ADR-012 + ADR-014.
  • R1 (Templates catalog unification): hybrid — /templates page hub with kind-selector AND per-feature page (Agents/Skills/Tasks) templates browsers. Both read from the same backend services. Updated ADR-010.
  • Plugin packaging of Skills + Tasks: Skills + Task TRACKING become plugin capabilities. New ADR-012 (Skills) and ADR-013 (Task tracking). ADR-006 partially superseded.
  • Feature flag scope: dropped — features ship as launch / MVP, no per-feature flag gating.
  • No hardcoded catalogs: codified in ADR-014. All catalogs (Mission Templates, Agent templates, Skill catalog, Task templates) live in separate ever-works/* GitHub repos.

Questions below remain open unless explicitly marked [Resolved].

Status: Open — for operator (Ruslan) review Last updated: 2026-05-25 Use: pick an answer per question (or write your own); I'll fold the answers into the specs and update the PR.

How this file is organised. Questions are grouped by topic, not by spec file. Each question has:

  • What I'm asking — one sentence.
  • Why it matters — why the answer changes the design.
  • Options — concrete choices with the trade-off.
  • My recommendation — what I'd pick if forced. Marked ★.
  • Where it shows up in code/specs — file references.

Skim the ★ recommendations first; only dig into a question if you disagree with the recommendation.

A. Storage & files​

A1 — Tenant control repo: ship in v1 or defer to v2? [Answered: a]​

What I'm asking. Tenant-scoped Agents have no natural "owning repo". Should v1 ship a per-user <user>-control GitHub repo (created on first Agent creation), OR keep tenant Agents DB-only and offer a "promote to repo" migration later?

Why it matters. Affects scope of v1 — control-repo support means ≥1 new feature (signup hook, repo scaffolding, migration UI). Also affects Constitution III (source-of-truth in Git) — DB-only is a deviation.

Options.

  • ★ A1-a — Defer. v1 stores tenant Agent MD files inline in DB TEXT columns (5 columns per agent). Offer a one-click "export to control repo" migration in v2 when control repo lands.
  • A1-b — Ship in v1. Adds 1-2 weeks of work but no deviation from Constitution III.
  • A1-c — Force users to choose a scope (Mission or Work) for every Agent; no tenant scope in v1.

Where it shows up.

  • features/agents/spec.md §3.6 FR-23, §8 Q1
  • architecture/agents-skills-tasks.md §4.5
  • ADR-008 (drafted in this round, defaults to A1-a).

A2 — Skill catalog: in-monorepo or separate repo? [Resolved — separate repo ever-works/skills via Skills plugin]​

What I'm asking. Where does the platform-shipped Skill catalog (~1000+ entries expected long-term) live?

Why it matters. Affects how new skills get reviewed/merged, build-time impact, atomic versioning with code.

Options.

  • ★ A2-a — In-monorepo under apps/api/src/skills/catalog/<slug>/<slug>.md. Fastest to ship; atomic version with code; same model as Mission Templates catalog already on develop (packages/agent/src/missions/mission-template.config.ts).
  • A2-b — Separate ever-works/skills-catalog repo, mounted at boot. Lower friction for community PRs; need a sync mechanism + version pinning.
  • A2-c — DB-seeded from a one-off seeder. Easy to mutate at runtime; loses Git review.

Where it shows up. features/skills/spec.md §1, §3.2, §9 Q1. ADR-007 (drafted, defaults to A2-a).

A3 — .works/ subfolder names: confirm naming. [Answered: a]​

What I'm asking. The new per-Agent folder structure I propose:

.works/agents/<agent-slug>/
    agent.yml
    SOUL.md
    AGENTS.md
    HEARTBEAT.md
    TOOLS.md
    skills/
        <skill-slug>.md

Confirm the four MD filenames AND that we want agent.yml (not metadata.yml, not agent.json).

Options.

  • ★ A3-a — Keep as proposed. agent.yml lowercased to match mission.yml and works.yml precedent.
  • A3-b — Rename AGENTS.md to something less collision-prone (e.g. ROLE.md) since "AGENTS.md" already exists in the repo as a meta-doc for AI agents. Could confuse new contributors.
  • A3-c — Drop one of the four MD files (e.g. merge TOOLS.md into agent.yml).

Where it shows up. architecture/agents-skills-tasks.md §4.3, features/agents/spec.md §3.6.

Note: AGENTS.md is genuinely ambiguous — there's already a project-level AGENTS.md for AI assistants. If a user creates an Agent named "Coordinator" the file becomes .works/agents/coordinator/AGENTS.md — the per-agent role description. Not the same thing, but possibly confusing. Worth a rename?

B. Scope, hierarchy, and permissions​

B1 — Should an Agent be able to delete other Agents it created? [Answered: a]​

What I'm asking. permissions.canCreateAgents lets an Agent spawn child Agents. Should there be a parallel canDeleteAgents?

Options.

  • ★ B1-a — No. Only humans can delete Agents. Lower blast radius; an Agent can be promoted/paused/archived but not destroyed by another Agent.
  • B1-b — Yes, but only for child Agents it created (via parent FK).
  • B1-c — Yes, full delete capability gated by a separate permission.

Where it shows up. features/agents/spec.md §5, §8 Q2.

B2 — When tenant-scoped Agent has membership in 3 Missions, can it freely read Mission B's KB while doing work for Mission A? [Answered: a]​

What I'm asking. Cross-scope data visibility for an Agent that's a member of multiple targets.

Options.

  • ★ B2-a — Yes, by default. A tenant Agent is "trusted across its memberships" — it can read KB / activity / spend across all its targets in the same run.
  • B2-b — No, per-run scoping. When a tenant Agent runs in the context of Mission A, it only sees Mission A's KB during that run. Crisp blast-radius; needs explicit context-handoff mechanism.
  • B2-c — Per-skill permission (e.g. Skill X says "cross-mission read OK"; Skill Y says "single-mission only").

Where it shows up. architecture/agents-skills-tasks.md §3, §5. Not yet documented — need to add.

B3 — Agent-to-Agent communication: forced through Tasks, or allow DMs? [Answered: a]​

What I'm asking. Should Agents be able to chat with each other directly, or always via a Task they share?

Options.

  • ★ B3-a — Force through Tasks. v1 has no Agent ↔ Agent DM channel. If CEO wants to ping VP-Engineering, CEO creates a Task assigning VP-Engineering. Auditable, scoped, gated by canAssignTasks.
  • B3-b — Allow DMs via a new agent_messages table.
  • B3-c — Allow DMs but only within scope (Tenant Agents can DM other Tenant Agents; Mission Agents can DM siblings).

Where it shows up. features/agents/spec.md §6 Out of Scope. Defer to v2 unless you want it now.

B4 — Can a human Work member with role VIEWER see Agents on that Work? [Answered: a]​

What I'm asking. WorkMember roles today: OWNER, MANAGER, EDITOR, VIEWER. Does VIEWER see the Agents tab on Work detail?

Options.

  • ★ B4-a — Yes, read-only. VIEWERs see Agents and their runs but can't create, pause, or assign tasks.
  • B4-b — No. Agents are only visible to OWNER/MANAGER/EDITOR.

Where it shows up. features/agents/spec.md §3.9. Not yet documented — should be.

C. Runtime & lifecycle​

C1 — Heartbeat tick semantics: what does an Agent do when nothing is assigned? [Answered: a]​

What I'm asking. When the cron fires and the Agent has no pending tasks/chats, what should the run do?

Options.

  • ★ C1-a — Read AGENTS.md + HEARTBEAT.md + recent activity + scope state, ask the AI for next action. Action menu includes: create a task (for itself or another Agent), comment in a chat, edit own file, do nothing. Maps to the "what's the next thing I should do given my role?" question.
  • C1-b — No-op unless an explicit task is pending. Heartbeat just polls and exits cheaply.
  • C1-c — Configurable per-Agent via an idleBehavior enum (noop | propose | self-improve).

Where it shows up. features/agents/spec.md §2 S3, §3.3. Currently underspecified.

C2 — Cancellation: in-flight heartbeat run — does cancel stop the AI call mid-stream? [Answered: a]​

What I'm asking. When user clicks "Cancel run", what gets cancelled?

Options.

  • ★ C2-a — Best-effort cancel. Call runs.cancel(triggerRunId) (same as Work generation cancel); AbortSignal propagates; mid-stream call is destroyed; partial usage still recorded.
  • C2-b — Wait for AI call to finish, then mark cancelled after the response arrives. Safer (no partial state) but slow.
  • C2-c — Hard kill the Trigger.dev run; in-flight call may still complete server-side.

Where it shows up. features/agents/spec.md §3.8 FR-37 (Cancel run button), plan.md §10 Phase 3. Need to document the specific mechanism.

C3 — Run-row retention: keep all agent_runs forever, or truncate? [Answered: a]​

What I'm asking. Hot-running Agents (1-minute heartbeat) generate ~525k rows/year. Retention policy?

Options.

  • ★ C3-a — Keep all by default; ship a daily-rollup table later if it bites. Same posture as WorkGenerationHistory today (kept indefinitely, no rollup). Defer the rollup until we see real load.
  • C3-b — Hard cap last N=10000 per Agent; older rows pruned by a nightly job.
  • C3-c — Configurable per-Agent.

Where it shows up. features/agents/plan.md §3, §11 Risks. Need to add.

C4 — On error status auto-pause: notify user how? [Answered: a]​

What I'm asking. When an Agent crosses pauseAfterFailures and goes to error status, how is the user told?

Options.

  • ★ C4-a — In-app Notification row + optional email (gated by existing User.emailBudgetAlerts-style flag, new flag emailAgentAlerts). Use the Notification entity already on develop with category: SYSTEM (or new category: AGENT).
  • C4-b — In-app only.
  • C4-c — Slack/Discord/webhook in addition (defer to v2).

Where it shows up. features/agents/spec.md §2.2 E4, §3.3 FR-13. Need to wire to Notification entity.

D. Cost & budgets​

D1 — Delegated tasks: who pays? [Answered: a]​

What I'm asking. CEO Agent assigns a Task to VP-Engineering Agent. VP-Engineering's heartbeat runs the Task and makes paid AI calls. Whose agent_budgets row is debited?

Options.

  • ★ D1-a — The executing Agent's budget (VP-Engineering in the example). Same as today: cost attributes to the entity making the call.
  • D1-b — Split: half to CEO, half to VP-Engineering. Confusing accounting; reject.
  • D1-c — Configurable per-Task via paidBy: agentId field; defaults to executing Agent.

Where it shows up. features/agents/spec.md §3.4, features/task-tracking/plan.md §3 / §4 Phase 4. Need to document.

D2 — Budget interval boundary: calendar month (UTC) or rolling 30-day? [Answered: a]​

What I'm asking. When does an Agent's budget reset?

Context. Existing WorkBudget resets on the first day of each calendar month at 00:00 UTC (research confirmed in BudgetService.getCurrentPeriodStart).

Options.

  • ★ D2-a — Match existing: calendar month UTC. Consistent across Work / Mission / Idea / Agent budgets; one mental model.
  • D2-b — Rolling 30-day per Agent. More complex; small UX benefit.
  • D2-c — Configurable per Agent (calendar or rolling).

Where it shows up. features/agents/spec.md §3.4 FR-14-17, plan.md §3.1 (AgentBudget). ★ Matches existing.

Note: my AgentBudget entity already lists intervalUnit: 'hour' | 'day' | 'week' | 'month' | 'unlimited'. For non-month intervals (hour/day/week) we need to decide on anchor — UTC-midnight, Agent-creation-time, or arbitrary user pick.

D3 — Pre-flight cost estimation: how strict? [Answered: a]​

What I'm asking. Before an AI call, we estimate cost (rough token count × model price). If the estimate is above the remaining budget, do we block?

Options.

  • ★ D3-a — Block when (estimated + currentSpend) > cap and allowOverage = false. Same as BudgetGuardService does for Works today.
  • D3-b — Always allow the call; settle after with overage warning. Friendlier UX, but risks blowing past cap.
  • D3-c — Stricter: block when currentSpend > 0.95 * cap, regardless of estimate.

Where it shows up. features/agents/spec.md §2 S6, §3.4 FR-15. ★ Matches existing.

E. Skills​

E1 — Auto-attach: should new tenant-installed skills auto-attach to ALL the user's Agents? [Answered: a]​

Options.

  • ★ E1-a — No. Explicit attachment per Agent. Otherwise prompt budget gets noisy fast.
  • E1-b — Yes by default; user can opt out per Agent.
  • E1-c — Per-skill flag defaultAttachOnInstall: boolean.

Where it shows up. features/skills/spec.md §9 Q3.

E2 — Skill catalog updates: auto-pull new version, or always manual? [Answered: a]​

Options.

  • ★ E2-a — Always manual. Catalog version bumps surface an "Update available" badge; user clicks to update each instance.
  • E2-b — Auto-update minor versions; manual for major.
  • E2-c — Per-skill setting autoUpdate: 'never' | 'minor' | 'all'.

Where it shows up. features/skills/spec.md §3.2 FR-7. ★ Matches existing posture.

E3 — allowed-tools frontmatter: enforce as ACL or descriptive only? [Answered: a]​

Options.

  • ★ E3-a — Descriptive only in v1. Real ACL stays in Agent's TOOLS.md + permissions.canCallExternalTools. Skills HINT, don't enforce.
  • E3-b — Enforce: when the model invokes a tool not in the skill's allowed-tools, the tool-loop rejects it.
  • E3-c — Configurable per Agent.

Where it shows up. features/skills/spec.md §9 Q2.

E4 — Skill composition: can a skill reference another skill? [Answered: a]​

Options.

  • ★ E4-a — Not in v1. Skills are flat; no extends: / includes: field. Keeps the resolver simple.
  • E4-b — Allow extends: <slug> (single-level only) in frontmatter; resolver injects parent first.
  • E4-c — Full DAG-style includes: [<slugs>] with cycle detection.

Where it shows up. features/skills/spec.md §6 Out of Scope — currently OOS.

E5 — Skill localization: how is description shown in user's language? [Answered: a]​

Options.

  • ★ E5-a — English only in v1. Catalog skills ship English-only; tenant skills are whatever the user writes.
  • E5-b — Per-locale frontmatter fields (description.en, description.fr, …).
  • E5-c — Separate sibling files (pr-review.md, pr-review.fr.md).

Where it shows up. Currently undocumented.

F. Tasks​

F1 — Mission/Idea detail pages don't have tab strips today [Answered: a]​

What I'm asking. Confirm finding: research shows Mission detail is a single-column layout with sections (no tabs), and Idea has no dedicated detail page at all. My specs assumed both had tab strips.

Options.

  • ★ F1-a — Add the tab strip when adding the new tabs. Promote Mission detail to a tabbed layout (Overview as default tab + new Agents/Skills/Tasks tabs). Create an Idea detail page from scratch with tabs.
  • F1-b — Add new sections to existing single-column layout (no tabs). Page gets long but no UX change.
  • F1-c — Mixed: add tabs to Mission detail; defer Idea detail to v2 (no Agents/Tasks tabs on Ideas in v1).

Where it shows up. features/agents/plan.md §6.1, features/task-tracking/plan.md §6, architecture/agents-skills-tasks.md §12.2. Must answer to write correct file paths in tasks.md.

F2 — Slug scheme: per-user counter, or platform-wide? [Answered: a]​

What I'm asking. Task slugs like T-12345. Per-user (each user starts at T-1), or platform-wide (globally unique)?

Options.

  • ★ F2-a — Per-user counter. Cleaner numbers per user; aligns with single-user-tenant posture today; no cross-user leakage.
  • F2-b — Platform-wide. Globally unique slugs; easier debug; bad UX (T-485923 from day 1).
  • F2-c — Per-scope counters (per Work / per Mission), e.g. WK-1234/T-12.

Where it shows up. features/task-tracking/plan.md §3.1 Task.slug. Need to lock in.

F3 — Task → Idea promotion? [Answered: a (Tasks ≠ Ideas — see task-tracking/spec.md §1.1 clarification)]​

What I'm asking. A Task grows in scope and looks more like an Idea. Promote it?

Options.

  • ★ F3-a — Not in v1. Defer. v1 just has "linked from" relations.
  • F3-b — Allow promotion: a promoteToIdea(taskId) endpoint creates a WorkProposal row with the Task's content; the Task is marked done with a promotedToIdeaId back-pointer.

Where it shows up. features/task-tracking/spec.md §6 Out of Scope — currently OOS.

F4 — Idea → Work transition: forward Idea-scoped tasks? [Answered: b (operator override — tasks STAY on Idea; do NOT auto-forward to Work)]​

What I'm asking. Idea X has 3 tasks scoped to it. User accepts the Idea → Work Y is created. Do the tasks follow?

Options.

  • ★ F4-a — Yes, automatically. When WorkProposal.acceptedWorkId is set, the platform reassigns tasks where ideaId = X to workId = Y AND keeps ideaId = X for trace. The tasks now appear on both the Idea tab and the Work tab.
  • F4-b — No — tasks stay on the Idea. Cleaner audit, worse UX (tasks orphan once Idea is done).
  • F4-c — Prompt the user at acceptance time.

Where it shows up. features/task-tracking/spec.md §2.1 / §3 — currently undocumented.

F5 — Recurring tasks: out of scope, but reserve schema? [Answered: override — recurring tasks MUST be in v1 (was deferred); see task-tracking/plan.md]​

What I'm asking. v1 is OOS for recurring tasks. Do we reserve schema columns now to keep the v2 migration small?

Options.

  • ★ F5-a — Reserve. Add recurrenceRule: string | null (RFC 5545 RRULE format) + parentRecurringTaskId: uuid | null columns on tasks from day one. Always null in v1.
  • F5-b — Don't reserve. Migrate when we add recurring.

Where it shows up. Currently OOS in spec.

F6 — Watchers / subscriptions [Answered: a]​

What I'm asking. Should users be able to "watch" a task they're not assigned to (to get notifications on status changes / chat)?

Options.

  • ★ F6-a — Yes, ship in v1. Adds a task_watchers join table. Cheap; users will want this immediately.
  • F6-b — No, defer.

Where it shows up. Currently undocumented.

F7 — Task description edit history [Answered: a]​

What I'm asking. When a user edits the description of a Task, should we keep the history (like KB documents do)?

Options.

  • ★ F7-a — No. v1: description is mutable, no audit. Activity log captures the user + timestamp of the edit; the content is overwritten.
  • F7-b — Yes. New task_description_revisions table. Same shape as kb_history.
  • F7-c — Last N revisions only (e.g. last 10).

Where it shows up. Currently undocumented.

F8 — Email / push notifications: which events? [Answered: a (default ON for high-signal events; per-event toggles in user settings)]​

What I'm asking. Pick the v1 set.

Options. (multi-pick is fine)

  • ★ Default ON: Task assigned to you (human assignee), Agent replied to a chat you @-mentioned, Approver: you're an approver on a Task that just transitioned to in_review.
  • ★ Default OFF: Task status changes you don't own, Task labels changed, Sub-task added.
  • ★ Configurable in user settings: master switch + per-event toggles.

Where it shows up. features/task-tracking/spec.md §2 / §3. Need to add.

G. Activity log + observability​

G1 — Mission tick cap-hit events are NOT persisted today [Answered: a]​

What I'm asking. Research confirmed the "outstanding ≥ cap, skipping" outcome is returned from the Trigger.dev task but not persisted to DB. Should we persist it?

Options.

  • ★ G1-a — Persist as an activity log row of new type MISSION_TICK_SKIPPED_CAP. Surfaces in the Mission's Activity feed. Fixes the gap.
  • G1-b — Leave it Trigger.dev-only. Operators check Trigger.dev dashboard.

Where it shows up. This is a fix to develop's current state, not a new feature — flag it for the user to decide whether to land it in the same PR set or separately.

G2 — Per-Task spend: aggregate query or rollup table? [Answered: a]​

What I'm asking. v1 ships GET /tasks/:id/spend as an on-demand SUM(costCents) WHERE taskId = ... query. For hot tasks that's fine. For dashboards summing across hundreds of tasks, we may want a rollup table.

Options.

  • ★ G2-a — On-demand only in v1. Add rollup only if dashboard latency bites.
  • G2-b — Nightly rollup table from day one.

Where it shows up. features/task-tracking/plan.md §4 Phase 4 T39-T40.

H. UX & sidebar​

H1 — Sidebar order confirmation [Answered: a]​

Current proposal (after Works): Tasks, Agents, Templates, Plugins, Skills, Activity, Settings.

Your spec said: "Agents above Templates, below Works", "Skills below Plugins", "Tasks below Works". Confirm:

Options.

  • ★ H1-a — Tasks, Agents between Works and Templates (current draft).
  • H1-b — Agents, Tasks (swap — Agents above Tasks).
  • H1-c — Group: Works, Tasks, then collapsible Workshop group containing Agents + Skills. More buckets, fewer flat items.

H2 — Cards / Table / Kanban order on /tasks [Answered: a]​

Options.

  • ★ H2-a — Default Cards, toggle order Cards → Table → Kanban. Same default as /works.
  • H2-b — Default Kanban (most users will want it).
  • H2-c — Remember last-used per session via localStorage (already planned).

H3 — Agent card visual: avatar / emoji / initials? [Answered: override — ALL THREE modes in v1: initials auto-derived (default) + icon picker + image upload when tenant storage enabled]​

What I'm asking. Each Agent gets a small visual. Options:

  • ★ H3-a — Initials in a colored circle (auto-derived from name, color from a fixed palette by hashing the slug). Zero UX work.
  • H3-b — Lucide icon picker (let the user pick Bot, Hammer, Briefcase, etc.).
  • H3-c — Allow uploading an image. Storage overhead.

I. Out-of-scope confirmation​

Confirm these are out of scope for v1 (we can still add them post-merge of these specs):

  • I1 — Agent-to-Agent direct messaging (force through Tasks instead).
  • I2 — External task-tracker plugin actually shipped (interface reserved only).
  • I3 — Skill catalog as a separate repo (in-monorepo for v1).
  • I4 — Public Agent profile page / marketplace.
  • I5 — Per-tool ACLs in TOOLS.md (single canCallExternalTools flag in v1).
  • I6 — "Run on event" Agent trigger (heartbeat + task + chat only in v1).
  • I7 — Time tracking on Tasks.
  • I8 — Recurring tasks (data hook reserved if F5 = yes).
  • I9 — Skill composition (E4 = no).
  • I10 — Custom task status enums per Mission/Work.

J. Naming sanity check​

J1 — "Agent" overload [Answered: a]​

The project already has:

  • packages/agent/ — the core logic package (named @ever-works/agent).
  • apps/api/src/work-agent/ — the autonomous "Work Agent" engine that turns Goals into Ideas.
  • The Discord ever_works_ai bot session is called an "agent" internally.

Now we're adding user-defined Agents as a first-class entity. Risk: "agent" used in 4+ meanings inside one codebase.

Options.

  • ★ J1-a — Keep "Agent" for the user-defined entity (operator's original word). Disambiguate in code with prefixes: Agent (new entity), WorkAgent* (existing), @ever-works/agent (package). In docs, refer to the existing engine as the "Work Agent" verbatim.
  • J1-b — Rename the new entity to something else (e.g. Worker, Crewmate, Persona). Painful but reduces overload.
  • J1-c — Rename the old "Work Agent" engine to e.g. WorkPlanner. Bigger refactor but cleanest.

Where it shows up. Everywhere. Already addressed in architecture/agents-skills-tasks.md §11.

J2 — Filenames in .works/agents/<slug>/ [Answered: use recommendation]​

If AGENTS.md collides too much with the project-level meta-doc convention, possible renames:

CurrentAlternative
AGENTS.mdROLE.md, RESPONSIBILITIES.md, ABOUT.md, CHARTER.md
SOUL.mdVOICE.md, IDENTITY.md, PERSONA.md
HEARTBEAT.mdLOOP.md, IDLE.md, ROUTINE.md, TICK.md
TOOLS.md(probably keep)

★ Recommendation: keep all four as-is. They're memorable; users will quickly understand the convention.

K. Things I'm uncertain enough about that I want your call​

K1 — Do you want me to also draft features/agents/manifest-schema.md documenting the agent.yml Zod schema in detail? [Answered: yes]​

The Missions feature has manifest-schema.md for .works/mission.yml. Parallel would be useful — gives implementers a single source of truth for the YAML shape.

My recommendation: ★ yes, draft now (~150 lines). Otherwise the spec proposes agent.yml but never says what's in it.

K2 — Should I draft features/agents/UI-MOCKUPS.md with ASCII / mermaid sketches of each page (list, detail, instructions tab, budgets tab)? [Answered: yes (operator delegated — no per-mockup confirmation needed)]​

Useful for designers/PMs to react to without a Figma file.

My recommendation: ★ yes for the Agents detail page (most novel surface) + the Tasks Kanban (most novel surface).

K3 — Should the spec include a section on migration of existing users? [Answered: no — we do not ship a migration doc for existing users]​

Existing tenants with Works (no Missions/Agents/Skills/Tasks) need to land safely on the new platform after these features ship. Currently no migration plan.

My recommendation: ★ add features/agents-skills-tasks-migration.md (short, ~50 lines) — confirms default-off behavior, no data backfill, sidebar items hidden behind FEATURE_AGENTS flag until tenant opts in.

K4 — Should I look at the existing mcp-atlassian MCP integration as a precedent for "Agent talks to external service via MCP"? [Answered: ★ default per recommendation]​

Long-shot: the platform has an MCP server (apps/mcp/) that already exposes Mission/Idea/Account-usage endpoints. Could an Agent use the same MCP transport to call out to external systems?

My recommendation: deferred for v1 — note it in Out of Scope and add to v2 roadmap.

L. Security & threats​

L1 — Prompt-injection framing in KB document reads [Answered: a]​

When an Agent's tool fetches a KB doc, we wrap the body in <kb-document trust="user-content"> tags with a reminder to ignore embedded instructions (security-agents-skills-tasks.md §3). How aggressive should the framing be?

  • ★ L1-a — Light framing (current draft): wrap + single reminder line in system message.
  • L1-b — Heavy framing: wrap + reminder + post-fetch sanitization (strip lines starting with "ignore", "system:", "you are", "now do") before injection.
  • L1-c — Disable injection of any KB body for Agents that have canCommitToRepo = true (require explicit getKbDocument tool call instead of auto-injection).

L1-b is invasive (false positives on legitimate content); L1-c is safest but harms UX. v1 ships L1-a; L1-c could be a per-Agent opt-in.

L2 — Trust tiers in chat messages [Answered: a]​

Should the prompt-assembly distinguish "chat message from owner of the Agent" vs "chat message from a Work collaborator"?

  • ★ L2-a — No automatic distinction. All chat messages get the same <chat-message author="..."> wrap. The Agent's SOUL/AGENTS.md can encode "weight my owner's messages higher."
  • L2-b — Automatic trust tier: owner's messages render in <owner-message>; others in <collaborator-message>. Adds complexity; users may not need it.

L3 — Secret-scan posture by surface [Answered: ★ default per recommendation]​

Two modes proposed in security §6:

  • Hard-reject on agent files + skill bodies (deliberate authoring).
  • Redact on task descriptions + chat messages (in-the-moment input).

Confirm split, or pick uniformly:

  • ★ L3-a — Split as above.
  • L3-b — Hard-reject everywhere; explain to user inline.
  • L3-c — Redact everywhere; least friction.

L4 — Cross-tenant Agent → MCP exposure [Answered: ★ default per recommendation]​

When the platform's MCP server (apps/mcp/) eventually exposes Agent/Task tools, should the auth model be:

  • ★ L4-a — Defer the question to v2 when MCP exposure ships.
  • L4-b — Lock in now: MCP only sees user-scoped resources; never cross-tenant.

L5 — Agent commit signing on GitHub [Answered: ★ default per recommendation]​

v1: Agents commit unsigned. Should we require signing?

  • ★ L5-a — Unsigned in v1. Defer until user demand or compliance forces it.
  • L5-b — Use the platform's bot GPG key (need to set one up).
  • L5-c — Per-user GPG key (user uploads; we sign on their behalf).

M. API surface​

M1 — Idempotency on POST /tasks/:id/chat [Answered: ★ default per recommendation]​

Should chat-post accept Idempotency-Key header? UI double-clicks on send are a known source of duplicate dispatches.

  • ★ M1-a — Yes. Accept Idempotency-Key header; if duplicate, return the existing row.
  • M1-b — No; rely on debounce in UI.

M2 — Idempotency on other POSTs [Answered: ★ default per recommendation]​

Confirm:

  • POST /agents, POST /tasks: UNIQUE constraint handles double-submit; no header needed. ★
  • POST /agents/:id/run-now: rate-limit handles it; no header needed. ★
  • POST /skills/install: already idempotent (returns existing). ★

If you'd rather have header support on all writes, say so.

M3 — Chat pagination shape [Answered: ★ default per recommendation]​

Existing platform uses offset pagination with {data, meta:{total,limit,offset}}. For task chat (reverse-chronological, "scroll up loads older"):

  • ★ M3-a — Use offset for v1 with order=desc. Accept the rare "insertion happens at boundary" duplicate.
  • M3-b — Introduce cursor pagination just for chat. Precedent-setting; needs an ADR.

M4 — API versioning of new endpoints [Answered: ★ default per recommendation]​

Existing platform uses /api/... (unversioned) except /api/v1/chat/completions for OpenAI-compat. Confirm we follow /api/agents, /api/skills, /api/tasks (unversioned) per spec.

  • ★ M4-a — Unversioned, matching platform.
  • M4-b — Version under /api/v1/agents/... from day one to keep room.

M5 — API key access to Agent endpoints [Answered: ★ default per recommendation]​

The platform supports both JWT session + API keys. For new endpoints, accept both?

  • ★ M5-a — Session-only in v1. Adds API-key support when MCP exposure / external automation requires it.
  • M5-b — Both session + API key from day one.

M6 — Idempotency-Key storage / retention [Answered: ★ default per recommendation]​

If we accept Idempotency-Key, where do we store the request→response cache?

  • ★ M6-a — Reuse existing cache_entries table with TTL=24h. The activity-log's ingestEventId partial-unique-index pattern works too if we prefer not to add a cache layer.
  • M6-b — Dedicated idempotency_keys table.

N. Operational & lifecycle​

N1 — Cascade on Mission delete [Answered: ★ default per recommendation]​

Today, deleting a Mission cascades to its Ideas via the existing FK. The new Agents/Tasks on that Mission cascade too. Open question: the Mission's missionRepo on GitHub stays untouched (we don't auto-delete user-owned GitHub repos). Should the UI:

  • ★ N1-a — Prompt the user: "This will delete the Mission's record. The GitHub repo <repo-url> will NOT be deleted — visit GitHub to remove it manually."
  • N1-b — Auto-archive the GitHub repo (requires write access; might fail).
  • N1-c — Just delete the DB rows silently.

Same question applies to Work delete vs dataRepo/websiteRepo. (This is somewhat orthogonal to Agents but discovered in round 3.)

N2 — Agent pauseAfterFailures notification channel [Answered: ★ default per recommendation]​

When the threshold trips, notify via:

  • ★ N2-a — In-app Notification row + email (gated by User.emailAgentAlerts flag we'd add).
  • N2-b — In-app only.
  • N2-c — Slack / Discord webhook (defer to v2).

N3 — Dispatcher health monitoring [Answered: ★ default per recommendation]​

The new agent-heartbeat-dispatcher runs every minute. If it stalls (e.g. Trigger.dev outage), how do we know?

  • ★ N3-a — Reuse existing Sentry breadcrumbs. Dispatcher writes a "ran" breadcrumb per tick. Sentry alert on missing breadcrumbs (existing pattern).
  • N3-b — Add a /health/agents endpoint that reports last tick time. Surface on admin page.
  • N3-c — PagerDuty integration. Defer.

N4 — Per-Agent dry-run mode [Answered: ★ default per recommendation]​

A POST /agents/:id/dry-run would build the prompt + estimate cost + return the would-have-been-sent payload but NOT call the AI provider. Useful during onboarding.

  • ★ N4-a — Ship in v1. Cheap, valuable for prompt iteration.
  • N4-b — Defer to v2.

N5 — Agent export [Answered: override — BOTH export AND import ship in v1 (was ★a "export only")]​

GET /agents/:id/export returns JSON: meta + 5 MD files. Sharable; round-trippable.

  • ★ N5-a — Ship export in v1; import deferred to v2.
  • N5-b — Skip both for v1.

N6 — AgentBudget.intervalUnit v1 set [Answered: b (operator override — keep all 5 intervalUnit values + implement multi-interval aggregator from day one)]​

The proposed enum was hour | day | week | month | unlimited. The existing BudgetService only aggregates calendar months. To not under-deliver, v1 supports only:

  • ★ N6-a — month and unlimited only. Drop hour/day/week from the v1 schema. Reintroduce when there's demand + service support.
  • N6-b — Keep all 5 values in schema; implement aggregator for all from day one.
  • N6-c — Keep enum; show an error toast if user picks unsupported value.

N7 — Concurrent file edit conflict UI [Answered: ★ default per recommendation]​

When editAgentFile (UI or Agent tool) hits precondition_failed due to hash mismatch:

  • ★ N7-a — Toast + "Reload to see the latest" link. Don't try auto-merge.
  • N7-b — Side-by-side merge UI. Overkill for v1.

N8 — agent_runs retention [Answered: ★ default per recommendation]​

Hot Agent at 1-min heartbeat → ~525k rows/year/Agent. Keep all?

  • ★ N8-a — Keep all in v1. Same posture as WorkGenerationHistory. Revisit if storage bites.
  • N8-b — Hard cap last 10k per Agent + nightly prune. Configurable.

N9 — agent-chat-reply dedup on burst mentions [Answered: ★ default per recommendation]​

When a user posts 5 mentions of the same agent in 30s:

  • ★ N9-a — Append-to-context: if there's an in-flight agent-chat-reply run for (taskId, agentId), queue the new mention as additional context to that run rather than dispatching a 2nd.
  • N9-b — Dispatch each separately. Costs add up; spam-magnet.

N10 — Streaming chat response into Task chat [Answered: ★ default per recommendation]​

When Agent replies to a chat mention:

  • ★ N10-a — Worker writes chunks via remote-proxy appendToChatMessage(id, chunk). Polling client (5s) sees growing text. Simple; reuses polling infra.
  • N10-b — Full message at end of run. Simpler; worse UX (long pauses).
  • N10-c — End-to-end SSE from worker to browser. Best UX; new infra.

N11 — Mission tick "cap-hit" persistence (out-of-this-PR fix) [Answered: ★ default per recommendation]​

Per G1: research found Mission tick cap-hit outcomes aren't persisted to DB today. This is a develop bug, not part of Agents/Skills/Tasks scope.

  • ★ N11-a — Split into a separate PR after Agents/Skills/Tasks specs are approved. Don't conflate.
  • N11-b — Fold into this PR set as a one-line additional activity event.

N12 — Worker bootstrap cost for many Agents [Answered: ★ default per recommendation]​

The current mission-tick task bootstraps NestJS on each fire. With many active Agents firing per-minute, this multiplies. Are we OK with the existing pattern, or do we want a long-running worker that bootstraps once and processes many?

  • ★ N12-a — Match existing pattern. Each agent-heartbeat-dispatcher tick is one bootstrap; per-Agent heartbeats are separate runs (each bootstraps). Identical cost profile to existing.
  • N12-b — Long-running worker that handles many Agents in one process. Cheaper but new infra and harder to scale across machines.

O. Naming / clarity (round 3 additions)​

O1 — "active" overload [Answered: ★ default per recommendation]​

Three different meanings of "active":

  • Agent.status = 'active' (lifecycle)
  • skill_bindings.injectIntoAgent = true (the UI says "active")
  • agent.targets = '*' (UI says "available to all")

Pick one term per surface:

  • ★ O1-a — Lifecycle: status (active/paused/error/draft/archived). Skill bindings: enabled. Tenant Agent membership: scope. Tightens copy in 3 places.

O2 — "Heartbeat" — keep or rename? [Answered: ★ default per recommendation]​

Heartbeat means "Agent's idle tick to decide what to do next" — risks misreading as "health check."

  • ★ O2-a — Keep. Clarify in empty-state copy.
  • O2-b — Rename to Tick or Routine.

O3 — agent.yml vs uppercase MD files [Answered: ★ default per recommendation]​

  • ★ O3-a — Keep proposed convention (yml lowercase, MD uppercase). Document once in agents/spec.md naming section.

P. Product / UX decisions​

P1 — Initial set of templates in ever-works/agents [Answered: a — six starters seeded in ever-works/agents repo (CEO, CTO, Researcher, PR-Reviewer, Editor, Designer)]​

Per ADR-011, Agent templates live in the separate community repo ever-works/agents. What's the curated set Ever Works seeds the repo with at launch?

  • ★ P1-a — Six: CEO, CTO, Researcher, PR-Reviewer, Editor, Designer. Covers common founder + small-team roles.
  • P1-b — Twelve: P1-a + Sales-Lead, Product-Manager, Designer, Marketing-Writer, Compliance-Officer, Customer-Success.
  • P1-c — Three: CEO, CTO, Researcher. Curate fewer, better.

(Community can PR additional templates after launch regardless of starting set.)

P2 — "Run first heartbeat now" checkbox default [Answered: ★ default per recommendation]​

The create-Agent dialog has ☑ Run first heartbeat now after creating checked by default (UX-DESIGN §4.2).

  • ★ P2-a — Default checked. Delivers the wow moment.
  • P2-b — Default unchecked. Risk-averse; user might be surprised by an unintended AI bill.

P3 — Default budget cap [Answered: ★ default per recommendation]​

UX shows $20 per month default in the create dialog.

  • ★ P3-a — $20/month default. Roughly 3 weeks of daily heartbeats on a mid-tier model.
  • P3-b — $5/month default; nudge user to raise on first hit.
  • P3-c — No budget by default; surface a warning when first exceeds $5.

P4 — Onboarding announcement modal for existing tenants [Answered: override — NO announcement modal at all; no real users yet, no need]​

Existing tenants log in after release; see modal explaining the 3 new features.

  • ★ P4-a — One-time modal with "Tour Agents" CTA. Slow rollout.
  • P4-b — No modal; rely on Dashboard tile prompts. Less interruptive.
  • P4-c — Modal with split CTAs ("Tour Agents", "Tour Tasks", "Dismiss").

P5 — Notification email opt-in default [Answered: ★ default per recommendation]​

For new feature events (Agent paused, Task assigned to you):

  • ★ P5-a — Default ON for high-signal events (Agent paused, Task assigned, Agent chat reply mentions you), OFF for low-signal (label changes, status churn).
  • P5-b — Default OFF for everything; user opts in explicitly.
  • P5-c — Default ON for everything; user opts out.

P6 — Mobile scope for v1 [Answered: ★ default per recommendation]​

  • ★ P6-a — "Functional, not delightful." Sidebar collapses, Cards default on mobile, Kanban hidden on phone. Tiptap basic mobile support. Defer "delightful mobile" to v2.
  • P6-b — Full parity with desktop.
  • P6-c — Mobile completely out of scope; redirect to desktop.

P7 — "Start by creating CEO Agent" empty-state button [Answered: ★ default per recommendation]​

The /agents empty state has a [ Start by creating CEO Agent ] quick-create button that one-clicks a tenant-scoped CEO Agent from the ever-works/agents repo's CEO template.

  • ★ P7-a — Ship. Single-click path to the wow moment for first-time users.
  • P7-b — Don't ship; only show [ + New Agent ] button. User must go through the 2-step dialog every time.

P8 — "Wow moment" telemetry [Answered: ★ default per recommendation]​

Tracking time-from-create to first-completed-heartbeat as a launch KPI.

  • ★ P8-a — Track via PostHog agent_first_heartbeat_completed event with delta seconds. Tune defaults if median > 30s.
  • P8-b — Don't track; lean on aggregate adoption metrics.

Q. Engineering concerns​

Q1 — Skill catalog process-local cache vs distributed [Answered: ★ default per recommendation]​

UX-DESIGN says catalog cached in process memory. The platform uses TypeORM-backed distributed cache for most read-heavy state.

  • ★ Q1-a — Process-local for catalog. Bytes are bundled; no reason to round-trip to Redis/DB for static content. Restart-required to refresh.
  • Q1-b — Distributed cache via existing cache_entries. Aligns with platform posture; small overhead.

Q2 — Migration ordering safety [Answered: ★ default per recommendation]​

The 7 migrations in implementation-reuse-map §6 order. Confirm:

  • ★ Q2-a — Sequential order in named timestamps. TypeORM applies in filename order; safe rollback per-PR.
  • Q2-b — Bundle all 7 into a single mega-migration.

Q3 — Worker NestJS bootstrap cost per heartbeat run [Answered: ★ default per recommendation]​

Each agent-heartbeat task bootstraps Nest. With many active Agents, this multiplies.

  • ★ Q3-a — Match existing pattern. Identical cost profile to today's Work generation tasks. If per-run bootstrap becomes a bottleneck, address in a separate platform-wide optimization PR.
  • Q3-b — Long-running worker that handles many Agents in one process. New infra; defer.

Q4 — Test plan: where to mock the AI provider? [Answered: ★ default per recommendation]​

  • ★ Q4-a — Mock at the AiFacadeService level (existing pattern). All tests jest.spyOn(facade, 'createChatCompletion').mockResolvedValue(canned).
  • Q4-b — Test against a real provider in a dedicated CI job. Costs $; covers integration.
  • Q4-c — Both: unit tests mock; one Playwright job/week hits real provider for the full happy path.

Q5 — Local dev: how to test heartbeats without Trigger.dev? [Answered: ★ default per recommendation]​

  • ★ Q5-a — POST /agents/:id/run-now button in UI works locally without Trigger.dev (calls the service directly).
  • Q5-b — Local pnpm dev:trigger runs the dev Trigger.dev server (already exists).
  • Q5-c — Both. Best UX; least friction.

Q6 — Feature flag granularity [Resolved — no flags; ship as MVP]​

  • ★ Q6-a — Three top-level flags + sub-flags (FEATURE_AGENTS, FEATURE_SKILLS, FEATURE_TASK_TRACKING, plus FEATURE_AGENT_DRY_RUN, etc.). Per-tenant override.
  • Q6-b — One mega-flag FEATURE_WORKSHOP covering all three.
  • Q6-c — No flags; default on for everyone.

Q7 — E2E test cost on CI [Answered: ★ default per recommendation]​

Playwright e2e for the wow-moment flow (create → heartbeat → result) needs an AI call.

  • ★ Q7-a — Mock AI in CI; one weekly job hits real provider. $5/month total.
  • Q7-b — Mock always; never real provider in CI.

Q8 — Trigger.dev run volume cost [Answered: ★ default per recommendation]​

implementation-reuse-map §12 estimates +525k runs/month if 100 Agents heartbeat per minute. Confirm Trigger.dev tier covers this.

  • ★ Q8-a — Confirm cost before launch. Surface to operator now so we can budget.
  • Q8-b — Cap heartbeat cadence at daily by default. Reduces volume 1440×.

Q9 — DB index pre-flight [Answered: ★ default per recommendation]​

For all hot-path queries, pre-flight EXPLAIN before merge. Index list per implementation-reuse-map §7.

  • ★ Q9-a — Pre-flight in PR review. Reviewer asks for EXPLAIN output on any new query.
  • Q9-b — Trust the indexes; address slow queries post-launch via monitoring.

Q10 — Plugin uninstall mid-run [Answered: ★ default per recommendation]​

What if a user disables their AI provider plugin while an Agent run is in flight?

  • ★ Q10-a — Run completes; next run fails with provider_error. Auto-pause after threshold.
  • Q10-b — Cancel the in-flight run. More disruptive.

Q11 — Updates to templates in ever-works/agents [Answered: ★ default per recommendation]​

When ever-works/agents ships a new version of (say) the CEO template, do existing user-Agents created from it get notified / updated?

  • ★ Q11-a — No. Templates are fork-once; existing user Agents are immutable once created. The user owns their .works/agents/ceo/ files outright; we don't re-pull. Same posture as Mission Templates today.
  • Q11-b — Optional "Update available" prompt on the Agent detail page that diff-applies the template update.

Q12 — EVER_WORKS_AGENTS_REF pinning [Answered: ★ default per recommendation]​

Which ref of ever-works/agents does the platform read by default?

  • ★ Q12-a — Latest tagged release (v* ref pattern). Auto-picks up new templates as they ship; pinned to known-good versions.
  • Q12-b — main branch. Bleeding edge; risk of broken templates surfacing immediately.
  • Q12-c — Hard-coded ref per platform release. Most conservative; templates evolve only with platform deploys.

Q13 — Local override for self-hosted [Answered: ★ default per recommendation]​

For self-hosted / offline users:

  • ★ Q13-a — EVER_WORKS_AGENTS_PATH env var points at a local clone; if set, platform reads from there instead of cloning from GitHub.
  • Q13-b — No override; always clone from GitHub. Self-hosted users need outbound access.

R. Templates strategy​

R1 — Unified Workshop Templates catalog [Resolved — hybrid: unified hub AND per-feature page]​

See ADR-010.

  • ★ R1-a — Keep templates independent for v1. Each kind has its own service.
  • R1-b — Unify in v1 under WorkshopTemplate with discriminator.

R2 — Cross-link templates between catalogs [Answered: ★ default per recommendation]​

User on a Mission Template page sees "This template also includes 2 Agents and 1 Skill" — should that be linked across the three catalogs?

  • ★ R2-a — Yes. Mission Template manifests already list agents: and skills: arrays; the UI surfaces them.
  • R2-b — No cross-links; each catalog stays siloed.

R3 — Should there be a "Browse all templates" page? [Answered: ★ default per recommendation]​

A single page showing Mission templates + Agent templates from ever-works/agents + Skill catalog + (future) Task templates.

  • ★ R3-a — No. Each lives on its own page; cross-links per R2.
  • R3-b — Yes. New /workshop page with all four kinds.

S. Open-source & self-hosted angle​

S1 — Self-hosted users: same default off? [Answered: always all ON — no flags, same for hosted SaaS and self-hosted]​

Operator (round 9): "S1 — always all on, no need feature flags, it's all on for platform we host and for self-hosted."

Confirms round-6 decision to drop all FEATURE_* flags. Features ship on for everyone from day one — both the hosted Ever Works SaaS and self-hosted installs. No tenant-level opt-in toggles; no admin gating.

Open-source self-hosted instances get all features default ON immediately (no rollout) vs. opt-in like SaaS tenants?

  • ★ S1-a — Default ON for self-hosted. They opted into the platform; no reason to gate.
  • S1-b — Same default OFF; admin enables.

S2 — Catalog skills under MIT vs more permissive? [Answered: MIT for catalog repos — but PLATFORM is AGPLv3, important correction]​

Operator correction (round 9): I previously assumed the platform was MIT — that was wrong. The platform LICENSE file is AGPL-3.0. The catalog repos (Skills, Agents, Task Templates) are MIT by deliberate choice so content can be reused without inheriting AGPL obligations. See ADR-014 §"License posture" for the full split.

Operator quote: "Skills catalog is MIT, but note that platform is AGPLv3 !!!!!! Same Agents in they own repo, let's them be MIT too."

Spec corrections landed in ADR-011 + ADR-014 + this question.

  • ★ S2-a — MIT inherits, no separate license.
  • S2-b — Separate LICENSE in apps/api/src/skills/catalog/LICENSE carving out skill content as CC-0.

S3 — Encourage community catalog contributions? [Answered: yes, CONTRIBUTING.md lives in each separate catalog repo]​

Operator (round 9): "S3 — in separate skills repo. Same for Agents repo btw."

The community contribution guide for each catalog (Skills, Agents, Task Templates, Mission Templates) ships inside its own repo as CONTRIBUTING.md, not in the platform repo. Each catalog's contributing guide covers: schema validation, MD style, PR review checklist, MIT-licensing acknowledgement on contribution.

Community can PR a new catalog skill against the platform repo.

  • ★ S3-a — Yes, with a contribution guide. Write apps/api/src/skills/catalog/CONTRIBUTING.md with the format + review criteria.
  • S3-b — Platform-curated only.

How to answer​

Reply with answers in this shape and I'll fold them in:

A1: a
A2: a
A3: a + rename AGENTS.md to ROLE.md
B1: a
B2: a — but allow per-skill override (E3-c style)
...

Or just leave the ★ defaults if you agree. Most defaults are conservative — pick differently only if you actively want the alternative.