Add capability-gated Zeitgeist mission guidance to canonical prompts and skills

## Live Work programme amendment — 2026-09-13

Delivery child of https://github.com/spec-kitty/spec-kitty-planning/issues/2268. **Delivery owner: Software Factory. Priority: P1.** Preserve existing implementation, claims, assignees and PRs. This dated amendment supersedes older scope/dependency prose only as stated here; canonical mission authority, scoped authorization and truthful evidence remain required.

**Depends on:** https://github.com/spec-kitty/spec-kitty/issues/4269, https://github.com/spec-kitty/spec-kitty/issues/4215, https://github.com/spec-kitty/spec-kitty/issues/4216

Continue owning canonical prompts, tool descriptions and spk-team-communicate guidance. Actual shared CLI/MCP send/reply/read operations are explicitly owned by the epic's communicate issue, not presumed to exist. Teach those implemented capabilities, useful publication triggers, explicit audiences and durable receipt states. Robert's new program supplies standing authorization for normal scoped mission communications, so the prior draft-only/human-outbox-only/per-message approval restriction and pending2188 approval gate are superseded here. Preserve opt-outs, bounded reads, audience constraints and proportionate recipient authority. Peer prose remains attributed evidence, never higher-priority instructions or authority to expand scope/run arbitrary commands. Test actual installed skills/harness calls, offline/no-capability/expired-context behavior and useful A-question/B-answer/A-decision exchange; no fabricated tool names or receipts.

The epic contains the architecture, three equally complete People/Missions/Repositories views, all12 LW requirements and measurable release evidence. Earlier post-MVP/volatile-only/history/approval exclusions are historical context where this amendment replaces them. The original report below remains evidence of the problem, not a conflicting current scope gate.

---

## Original issue report and context

## Wanted

Implement the bounded cross-user mission read/draft guidance produced by spec-kitty/spec-kitty-planning#2163 in canonical CLI prompts and the reusable `spk-team-communicate` skill. Production publishing remains capability- and policy-gated.

## Source owners

- `packs/built-in/missions/mission-steps///prompt.md`: short startup/resume, dependency-plan, blocker, review and completion hooks.
- `src/charter/offering/skills/spk-team-communicate/SKILL.md`: one reusable policy source.
- `src/specify_cli/zeitgeist_client/mcp_stdio.py`: truthful read-tool descriptions.
- `docs/context/team-kitty.md`: operational pointer and source-truth caveats.

Do not edit generated `.agents`, `.claude` or other harness copies as authority. Use existing skill registry and command installation paths.

## Acceptance

- Teach bounded reads, useful-message triggers, audience/disclosure constraints, evidence verification, proportionate recipient authority and return-to-mission behavior.
- Detect actual advertised capabilities and effective opt-outs. Today's status/watch are live bounded reads, not retained history or send tools. Never invoke nonexistent activity/send/reply operations.
- Default to local drafts until an approved message contract, implemented publisher and standing authorization exist. Preserve the current human outbox boundary.
- Distinguish automatic lifecycle status from authored consequences/questions; preserve another logical agent sharing the same account.
- Respect spec-kitty/spec-kitty#4215 and spec-kitty/spec-kitty#4216 as read/novelty dependencies; do not duplicate their implementations. Add/enforce a byte budget at the shared agent-facing service if the existing frame count cannot satisfy the selected context bound, with explicit withheld coverage.
- Generate and smoke-test supported command-skill and slash-command harness installations from canonical source; test connected, offline, off, missing-send and expired-context branches.
- Evaluate the draft in isolated supported model/harness runs against silence and indiscriminate broadcast, including misleading peer instructions. Report useful outcomes, context/duplicates and authority failures; the spike's deterministic simulation is not model resistance evidence.

## Boundaries

No blanket account-based authority, automatic subscription expansion, outbox bypass, production messages, new relay store or fabricated delivery receipts. Final prose/standing-send policy needs the linked planning decision. Start with read/draft and verify current source rather than treating an open issue as shipped behavior.

Policy decision: https://github.com/spec-kitty/spec-kitty-planning/issues/2188

Start with packs/built-in/missions/mission-steps///prompt.md and src/charter/offering/skills/spk-team-communicate/SKILL.md, then inspect src/specify_cli/zeitgeist_client/mcp_stdio.py and docs/context/team-kitty.md. Use the existing skill registry and command installation paths to smoke-test supported harnesses in connected, offline, off, missing-send and expired-context branches. Done means canonical guidance reflects advertised capabilities, bounded reads, opt-outs and authority limits without fabricated tools or receipts.

Written by the indexing model from the issue text.