wa-prompt-only-sanctuary-dialog-opportunities

Imported from jleechanorg/jleechanbrain (skills/wa-prompt-only-sanctuary-dialog-opportunities/SKILL.md). Install upstream with npx skills add jleechanorg/jleechanbrain --skill wa-prompt-only-sanctuary-dialog-opportunities. Copyright stays with the author (MIT).

Add a prompt-only "BG3-style companion dialog" rule to a WA campaign

When to use

User asks for a dialog-only, rest-anchored, no-plot-advancement scene surface (BG3-style campfire/lunch/dinner banter with companions) inside an existing campaign. The right fix is almost always a prompt-only edit that anchors on an existing mechanical state flag (e.g. custom_campaign_state.sanctuary_mode.active) — not a new schema field, not a new agent, not a backend state change.

Concrete trigger phrases: "BG3-style companion dialog", "dialog opportunities during long rest / short rest / lunch / dinner / breakfast", "companion banter at campfire", "sanctuary-mode quiet scene", "no plot advancement", "rest-anchored beat".

Why this is a prompt-only change

The user's locked preference (campaign-design-rpg-bible rule #2) is to never pre-decide outcomes, always offer a 3-option carousel + Freeform slot. The existing Sanctuary Mode section in mvp_site/prompts/living_world_instruction.md already says:

ALLOWED during sanctuary: Companion conversations, planning, shopping, training, peaceful exploration, minor (non-lethal) complications.

…but that's a passive allowance. The LLM needs a concrete, repeatable rule. Add a self-contained section that lifts the passive allowance into an active trigger + a worked example + a hard-prohibitions list.

The 4-shaped rule (always emit all four)

1. When to offer — gate on the existing state flag + a list of rest-anchored trigger verbs (long rest, short rest, lunch, dinner, breakfast, campfire, evening/morning watch, "we make camp", "we eat", "we sit by the fire").
2. What the opportunity looks like — cast (1 companion by default, 2 if recent turns featured both), setting (anchor physically to where the player actually is), dialogue (reference concrete companion_arcs.<name>.history[last] or core_memories), 3-option tonal carousel + Freeform.
3. What the opportunity MUST NOT do — no plot advancement, no companion arc progression, no scene events, no XP, no rest_taken, no new god-mode directives, world time advances < 1 hour.
4. Output shape — ride on the existing character-mode narrative response; no new top-level field; dialog in narrative, choices in planning_block.choices.

Where to anchor the section

Append the new section inside the existing Sanctuary Mode / sanctuary section of mvp_site/prompts/living_world_instruction.md, immediately after the existing Sanctuary Rules block (right before ## Output Requirements). Do NOT create a new file. Do NOT add to PATH_MAP — the file is already registered.

The contract test (mandatory per mvp_site/prompts/AGENTS.md Hard rule #5)

Add a single test to mvp_site/tests/test_prompts.py that pins:

• the section anchor (## Sanctuary Dialog Opportunities)
• the eight trigger verbs (assert each one in the loop)
• the Freeform slot literal ("Say something else (freeform)")
• the companion_arcs anchor
• the six hard prohibitions ("No plot advancement", "No companion arc progression", "No scene events", "No XP", "rest_taken", "companion_arc_event`)
• the output shape (narrative + planning_block.choices)

Use the existing _load_instruction_file / _read_prompt helper that resolves {{PROMPT_INCLUDE:}} placeholders — never read the raw file in tests.

Hard rules (from mvp_site/prompts/AGENTS.md)

1. No version history in the prompt body.
2. No meta-commentary / "this is the most important section" / "remember to do X".
3. No editorializing.
4. Each rule self-contained in the served prompt (don't share via silent copy).
5. No banned entities (no D&D Ao/Mystra/Shar unless the prompt is explicitly the D&D appendix).
6. Generic placeholders, then setting-specific appendix.
7. No code, no commands, no shell in prompt text.
8. Quantitative anchors with formulas, not inspirational cites.
9. Write for the LLM that executes, not for a human reviewer.

Workflow (track all of these)

1. Worktree from origin/main — git worktree add -b feat/<slug> ~/.worktrees/wa-<slug> origin/main. Never push onto a feature branch.
2. Symlink the venv — ln -s ${HOME}/projects/worldarchitect.ai/venv venv so ./vpython works in the worktree.
3. Edit the prompt — single self-contained section (see "The 4-shaped rule").
4. Add the contract test — see "The contract test".
5. Run the test — ./vpython -m unittest mvp_site.tests.test_prompts.TestPromptLoading.test_<name> -v and the surrounding sanctuary/prompt-registration tests. PASS required.
6. Verify the served prompt — python -c "from mvp_site.agent_prompts import _load_instruction_file; p = _load_instruction_file('living_world'); assert 'Sanctuary Dialog Opportunities' in p; print('OK')" to confirm the rule reaches the LLM.
7. File the bead — br create "<user-ask title>" --type chore --priority 2 --description-file <path> for the work item.
8. Commit + push — commit with <cli>/<model>: prefix per SLT rule; push the branch.
9. Open the PR — gh pr create with the worked example inside the body (per AGENTS.md requirement).
10. Slack the user at channel root (not thread — mobile doesn't render threads) with prompt + PR + contract test + bead.
11. Arm the one-time status cron — cronjob create "20m" with --delete-after-run, deliver to the same channel.

WORKFLOW CORRECTION (skillify lesson from 2026-08-17, PR #8979)

/ready is a CI gate (Green Gate + no conflicts). It does NOT fetch inline review comments. After creating a PR, treat the FIRST 60 minutes as a recent-commit-recheck window — on every Slack-side re-entry to the PR, run:

gh api graphql -f query='{ repository(owner: "jleechanorg", name: "worldarchitect.ai") {
  pullRequest(number: N) {
    reviewThreads(first: 50) { nodes { id isResolved comments(first: 5) { nodes { author { login } body } } } }
  }
}}'

If any isResolved: false thread is from the user, address it before posting any "ready for review" / "all green" reply. Reply to thread via addPullRequestReviewThreadReply(input: {pullRequestReviewThreadId, body}); resolve via resolveReviewThread(input: {threadId}). Don't trust the visible PR state (isDraft:false, mergeable:MERGEABLE) — review threads are orthogonal.

Pitfalls

• Don't create a new top-level field. Sanctuary dialog opportunities ride on the existing narrative + planning_block.choices pair. Adding a new field forks the schema and the rendering pipeline.
• Don't make it a companion_arc_event. That system is for multi-turn arc progression (20-30 turns from discovery to resolution). A BG3-style campfire is a beat, not a step. The hard-prohibition list explicitly forbids companion_arcs.<name>.progress / phase / history writes.
• Don't emit rest_taken from a dialog opportunity. Mechanical rest is the player's separate choice; emitting rest_taken here would reset resource registries and front-load the rest side-effects into the dialog turn.
• Don't advance world_data.world_time by days. The opportunity is a beat that lasts a few minutes; the player can take a real mechanical rest on the next turn.
• Don't pick the same companion two offers in a row unless the player asked. Read companion_arcs.<name>.history and pick the one with the least-resolved recent topic.
• Don't forget the Freeform slot. The user's locked preference (campaign-design-rpg-bible rule #2) is a 3-option carousel + Freeform in every choice surface. Forgetting Freeform is the single most common LLM regression on this campaign class.
• Don't claim it's "done" without the contract test passing. Per mvp_site/prompts/AGENTS.md Hard rule #5, the test is the canonical evidence the rule is in the served prompt.

Verification

• New section appears in _load_instruction_file('living_world') output (assert 'Sanctuary Dialog Opportunities' in p)
• Contract test passes (4/4 incl. surrounding sanctuary + prompt-registration tests)
• No new top-level field in any schema
• No PATH_MAP change (the file is already registered for StoryModeAgent)
• PR is clean off origin/main (git log origin/main..HEAD shows only this commit)
• Diff budget < 1000 lines (typical: +54 prompt, +40 test)
• Bead filed at P2
• Slack reply at channel root
• One-time status cron armed (--delete-after-run, --at 20m)

Cross-references

• ~/.smartclaw/skills/campaign-design-rpg-bible/SKILL.md rule #2 (3-option carousel + Freeform)
• ~/projects/worldarchitect.ai/mvp_site/prompts/AGENTS.md Hard rules 1-9 (prompt content contract)
• ~/projects/worldarchitect.ai/mvp_site/prompts/living_world_instruction.md § Sanctuary Mode (existing mechanical sanctuary)
• ~/projects/worldarchitect.ai/mvp_site/prompts/living_world_instruction.md § Companion Quest Arcs (existing multi-turn arc system — explicitly NOT triggered)
• ~/projects/worldarchitect.ai/mvp_site/tests/test_prompts.py:test_rewards_prompt_carries_sanctuary_activation_contract (precedent for the contract test pattern)
• SOUL.md ## COMMIT: always-skillify-after-non-trivial-work (this skill is the durable artifact)
• SOUL.md ## COMMIT: one-time-status-cron-after-every-task (the 20m follow-up cron)
• SOUL.md ## COMMIT: pr-clean-branch-from-main-no-history-bloat (clean-from-origin/main worktree pattern)