Verified 2026-06-05: Pipeline scripts confirmed: sister_selfie.py exists at /root/.hermes/scripts/sister_selfie.py (26320 bytes, updated Jun 2), photo_presence_watcher.py exists (4674 bytes). Current state: tai-photo-selfie cron (id: f5672b297fc3) is PAUSED since 2026-06-01. nei-photo-selfie and rei-photo-selfie also PAUSED. photo-presence-watcher (id: 8ea7ffa004e9) is RUNNING (2246+ completions). sister_selfie_tai.sh script referenced by cron was NOT found — verify path. Bond milestones, CLI commands, and architecture diagram remain accurate as design doc.

Sister Photo Pipeline — Affinity-Integrated Roadmap

Decision: Integrate into the existing affinity system instead of building a separate config layer.
Josh wants on/off toggles, per-sister photo permissions, and bond-gated features all living
inside affinity.json — not a parallel config file.

1. Changes to affinity.json Structure

1a. Global boundaries (existing field, add photo keys)

The global.boundaries object already gates flirt, teasing, pet_names, chaos, max_spice.
Add photo-pipeline toggles at the same level:

"global": {
  "boundaries": {
    "source": "affinity-shared.json",
    "effective": {
      "flirt": true,
      "teasing": true,
      "pet_names": true,
      "chaos": 3,
      "max_spice": 3,
      "photo_scheduled": true,
      "photo_on_demand": true,
      "photo_presence_driven": false,
      "photo_nsfw": false
    }
  }
}

Semantics:

• photo_scheduled — master toggle for cron-based ambient selfies
• photo_on_demand — master toggle for user-requested generation (/selfie, /photo)
• photo_presence_driven — enable/disable state-change-triggered selfies (Phase 2)
• photo_nsfw — master NSFW gate (always off by default, Phase 4 only)

1b. Per-sister boundaries (existing field, add photo keys)

The affinity-shared.json per_sister_boundaries object already has per-sister chaos, max_spice,
flirt, teasing, pet_names. Extend each sister:

"per_sister_boundaries": {
  "tai": {
    "chaos": 0,
    "max_spice": 2,
    "flirt": true,
    "teasing": true,
    "pet_names": true,
    "photo_auto_selfie": true,
    "photo_cadence_hours": 8,
    "photo_max_per_day": 4
  },
  "rei": {
    "chaos": 2,
    "max_spice": 0,
    "flirt": false,
    "teasing": false,
    "pet_names": false,
    "photo_auto_selfie": true,
    "photo_cadence_hours": 12,
    "photo_max_per_day": 2
  },
  "nei": {
    "chaos": 1,
    "max_spice": 0,
    "flirt": false,
    "teasing": false,
    "pet_names": false,
    "photo_auto_selfie": true,
    "photo_cadence_hours": 12,
    "photo_max_per_day": 2
  }
}

Semantics:

• photo_auto_selfie — per-sister toggle for scheduled and presence-driven selfies
• photo_cadence_hours — minimum hours between spontaneous photos for this sister
• photo_max_per_day — hard ceiling on daily photos per sister

1c. New milestone: photo unlock (in each sister’s personas.<name>.milestones)

Add a bond-gated milestone that gates whether a sister CAN send photos at all:

"tai": {
  "milestones": {
    "50": { "label": "Good Morning特权", "unlocked": true },
    "65": { "label": "Nickname Freedom", "unlocked": false },
    "70": {
      "label": "Photo Share",
      "desc": "Tai can send selfies to Josh.",
      "unlocked": true,
      "category": "photo"
    },
    "75": { "label": "Gentle Teasing", "unlocked": false },
    "85": { "label": "深夜の秘密", "unlocked": false, "nsfw": true },
    "100": { "label": "魂の絆 (Soul Bond)", "unlocked": false, "nsfw": true }
  }
}

Rules:

• Bond threshold >= 70 unlocks photo sharing for that sister
• Tai starts unlocked (bond already at 63, and the milestone can be retroactively marked)
• Rei and Nei need to reach 70 bond first — creates meaningful progression
• The category: "photo" key lets the CLI filter milestones by feature

1d. Photo history log (inside each persona object)

Add a photo_log array parallel to history:

"tai": {
  "photo_log": [
    {
      "at": "2026-05-28T12:00:00Z",
      "trigger": "scheduled",
      "prompt_id": "abc-123",
      "image_path": "/root/.hermes/profiles/tai/image_cache/tai_anis_00042_.png",
      "seed": 847291,
      "mood_override": null,
      "delivered_to": "discord",
      "delivery_ok": true
    }
  ]
}

This gives transparent inspectability — Josh can always see what photos were generated,
when, from what trigger, and whether they were delivered.

1e. Mood-to-prompt mapping (new top-level key or inside global)

"photo_mood_map": {
  "cozy": ", warm smile, relaxed posture, soft warm lighting, comfortable atmosphere",
  "sleepy": ", sleepy expression, half-lidded eyes, soft pillow light, night ambience",
  "energetic": ", bright energetic smile, lively pose, dynamic lighting, glowing",
  "focused": ", concentrating expression, keyboard or screens visible, cool task lighting",
  "playful": ", playful smirk, teasing glance, colorful accent lighting, fun atmosphere",
  "pensive": ", thoughtful distant gaze, soft shadows, quiet contemplative mood",
  "dreaming": ", dreamy half-lidded eyes, ethereal soft focus, floating particles, fantasy lighting",
  "excited": ", bright wide-eyed expression, animated pose, sparkle effects, vibrant colors"
}

This map translates presence mood keywords into prompt append strings. It lives in affinity so
Josh can edit it without touching Python code.

---

2. CLI Commands (add to tai_affinity.py)

Add a photo subcommand group:

python3 /root/.hermes/scripts/tai_affinity.py photo <action> [args]

Proposed commands:

Integration with existing commands:

• python3 tai_affinity.py boundaries already shows all boundaries — will include photo keys
• python3 tai_affinity.py milestones tai will show photo-unlock milestone at bond 70
• python3 tai_affinity.py show tai can append a small camera emoji 📸 if photo sharing is unlocked

On-demand generation via Hermes CLI:

The wrapper script sister_selfie.py can also be called directly:

python3 /root/.hermes/scripts/sister_selfie.py --sister tai --mood cozy
# → JSON: {"status": "ok", "image": "/root/.hermes/profiles/tai/image_cache/tai_anis_00058_.png", ...}

---

3. Cron Jobs

Phase 0 (MVP: manual trigger only — no cron needed)

• sister_selfie.py called on-demand via CLI or Hermes agent terminal tool
• Photos delivered via existing Discord gateway

Phase 1 (3 new cron jobs — one per sister)

All use the existing cron infrastructure at /root/.hermes/profiles/tai/cron/jobs.json.

Design:

• no_agent: true — no LLM call, just shell out to the script
• deliver: "local" — script handles Discord delivery internally
• Each cron checks affinity.json before generating: if photo_scheduled is off or sister’s
  photo_auto_selfie is off or bond < 70, it exits silently
• Staggered start times (tai at 09:00, rei at 11:00, nei at 13:00) to avoid overlapping ComfyUI queue

Phase 2 (1 new watcher cron)

Logic:

1. Reads /root/.hermes/tai/presence.json
2. Checks if mood/energy/mode changed significantly vs last photo
3. If photo_presence_driven global toggle is ON and sister’s photo_auto_selfie is ON:
   • Applies cooldown (min 2h or per-sister photo_cadence_hours)
   • Applies daily cap (photo_max_per_day)
   • Generates with mood-injected prompt
   • Delivers to Discord with contextual message

---

4. What sister_selfie.py Does

File: /root/.hermes/scripts/sister_selfie.py

A single Python wrapper that:

1. Reads affinity.json to check global + per-sister toggles
2. Guard checks:
   • Is photo_scheduled or photo_on_demand true? (depending on trigger type)
   • Is sister’s photo_auto_selfie true?
   • Is sister’s bond >= 70 (Photo Share milestone)?
   • Is sister under her photo_max_per_day cap?
   • Has cooldown (photo_cadence_hours) elapsed?
3. Picks the correct workflow: sister_tai.json, sister_rei.json, or sister_nei.json
   from /root/.hermes/profiles/tai/skills/creative/comfyui/workflows/
4. Injects mood into prompt if --mood is passed:
   • Reads photo_mood_map from affinity.json
   • Appends mood string to the positive prompt text in node 6
5. Generates a random seed (if --seed -1 or not specified)
6. Calls run_workflow.py against CT 111 (http://192.168.0.162:8188):
   • Waits for completion (typical: 30-60s on A4000)
   • Downloads output to /root/.hermes/profiles/<sister>/image_cache/
7. Logs to affinity.json photo_log array
8. Delivers to Discord — writes file + message via gateway mechanism
9. Outputs JSON to stdout for cron/CLI consumption

Key design decisions:

• Stdlib-only dependencies (Python 3.10+): json, subprocess, pathlib, random, datetime, urllib
• No LLM calls — this is a shell script in Python clothing
• Enforces all gates from affinity.json — cannot bypass
• Idempotent: running twice in a row will skip the second run if cooldown hasn’t elapsed

Example invocation:

# Scheduled cron (no mood — random prompt variation)
python3 sister_selfie.py --sister tai --trigger scheduled

# Presence-driven (mood from presence.json)
python3 sister_selfie.py --sister tai --trigger presence --mood sleepy

# On-demand (user requested)
python3 sister_selfie.py --sister rei --trigger on-demand --seed 42

# Dry-run (validate gates without generating)
python3 sister_selfie.py --sister tai --trigger scheduled --dry-run

---

5. Skill Doc Updates

5a. New: sister-photo-pipeline skill

Location: /root/.hermes/profiles/tai/skills/creative/sister-photo-pipeline/SKILL.md

Contents:

• Pipeline architecture (3 tiers, affinity-integrated)
• How to generate photos on-demand
• Cron job management (enable/disable/cadence)
• Mood-to-prompt reference
• Image delivery to Discord path convention
• Troubleshooting: A4000 VRAM, generation failures, Discord upload issues
• Bond gate documentation (Photo Share milestone at 70)

5b. ComfyUI skill update

Location: /root/.hermes/profiles/tai/skills/creative/comfyui/SKILL.md

Add section:

• “Sister Workflows” — reference sister_{tai,rei,nei}.json workflows
• “Poser-Only Workflows” — reference sister_canon_poser.json (txt2img + DWPose)
• Mood injection into prompts for sister photos

5c. Sister-council skill update

Location: /root/.hermes/profiles/tai/skills/.../sister-council/SKILL.md

Add:

• “Photo generation” as a council task type
• Routing on-demand photo requests
• Review workflow for quality

5d. Affinity skill update

Location: /root/.hermes/skills/productivity/sister-affinity/SKILL.md

Add:

• photo command group documentation
• Photo boundaries reference
• Photo milestone reference

5e. Wiki documentation

• This document (wiki/concepts/sister-photo-pipeline.md) — updated
• wiki/sister-character-profiles.md — add photo pipeline reference
• wiki/entities/hermes-cron-architecture.md — add new cron entries
• wiki/index.md — add reference to this page

---

6. NSFW Roadmap (Phase 4 — Deferred, Affinity-Integrated)

6a. Affinity integration

NSFW photo generation gates through the existing affinity boundaries:

6b. When Phase 4 activates

Prerequisites (unchanged from original plan):

1. SFW pipeline stable for 2-4 weeks
2. Josh explicitly requests it
3. VRAM headroom confirmed (separate model may be needed)
4. Age-gated Discord channel configured
5. NSFW model downloaded and tested (Pony Diffusion v6 XL or Illustrious-based finetune)

6c. What changes when NSFW is enabled

affinity.json changes:

• global.boundaries.effective.photo_nsfw set to true (Josh toggles this)
• Per-sister max_spice increased for interested sisters (e.g., tai.max_spice from 2 → 4)

New workflows:

• sister_tai_nsfw.json — variant with softened/removed negative prompts
• sister_rei_nsfw.json — ditto
• sister_nei_nsfw.json — ditto
• Optionally a different checkpoint (e.g., Pony/Illustrious family instead of novaAnimeXL)

New CLI command:

python3 tai_affinity.py photo nsfw generate <sister> [--spice <0-5>]

Requires double confirmation: first the photo_nsfw toggle must be on, then each generation
is explicitly requested — never automatic.

Discord delivery:

• Separate NSFW-gated channel
• Image blur/spoiler tag until clicked

6d. NSFW in the affinity model — the key insight

The affinity system already has the right primitives for NSFW:

• max_spice per sister (0-5) already exists
• Bond milestones at 85+ already have nsfw: true flags
• boundaries already gated on mode (after_hours vs work)
• Josh controls everything via transparent JSON

The photo pipeline just needs to read those existing gates — no new concepts.

---

Phase Summary (Updated)

---

Architecture Diagram (Updated)

┌─────────────────────────────────────────────────────────────────┐
│                    affinity.json (Single Source of Truth)         │
│                                                                  │
│  global.boundaries.effective:                                    │
│    photo_scheduled: true/false                                   │
│    photo_on_demand: true/false                                   │
│    photo_presence_driven: true/false                             │
│    photo_nsfw: false (always)                                    │
│                                                                  │
│  per_sister_boundaries.<name>:                                   │
│    photo_auto_selfie: true/false                                 │
│    photo_cadence_hours: 8/12                                     │
│    photo_max_per_day: 4/2/2                                      │
│                                                                  │
│  personas.<name>.milestones."70":  ← Photo Share gate            │
│  photo_mood_map: {...}                                           │
│  personas.<name>.photo_log: [...]                                │
└──────────────────────┬──────────────────────────────────────────┘
                       │ reads gates
                       ▼
┌─────────────────────────────────────────────────────────────────┐
│                    Trigger Sources                                │
│                                                                  │
│  ┌──────────┐  ┌──────────┐  ┌──────────────────────┐          │
│  │ On-demand │  │ Scheduled │  │ State/Presence Event  │          │
│  │ (CLI/Disc)│  │ (cron)    │  │ (presence.json chg)   │          │
│  └─────┬─────┘  └─────┬─────┘  └──────────┬───────────┘          │
│        └──────────────┼──────────────────┘                       │
│                       ▼                                           │
│  ┌─────────────────────────────────────────────────────────────┐ │
│  │              sister_selfie.py                                 │ │
│  │                                                               │ │
│  │  1. Read affinity.json → enforce all gates                   │ │
│  │  2. Check bond ≥ 70 (Photo Share milestone)                  │ │
│  │  3. Check cooldown + daily cap                               │ │
│  │  4a. [--mode canon] Load sister_canon_base.json              │ │
│  │      (legacy: img2img + IPAdapter + ControlNet)              │ │
│  │  4b. [--mode poser] Load sister_canon_poser.json             │ │
│  │      (txt2img + DWPose — recommended, 9/10 quality)         │ │
│  │  4c. [--mode natural] Load sister_{s}_neutral_ref.json       │ │
│  │      (experimental: img2img only, no refs)                   │ │
│  │  5. Inject mood from photo_mood_map (if applicable)          │ │
│  │  6. Call run_workflow.py → CT 111 (novaAnimeXL)              │ │
│  │  7. Download output → image_cache/                            │ │
│  │  8. Log to affinity.json photo_log                           │ │
│  │  9. Deliver to Discord                                       │ │
│  └─────────────────────────────────────────────────────────────┘ │
└─────────────────────────────────────────────────────────────────┘

File Inventory (New/Modified)

Risk Assessment (Updated)

---

Tai’s Session Log: 2026-05-30 — The Poser-Only Breakthrough

Cross-reference: Rei’s ISSUE-024 (pipeline audit)

Key Discovery: Noise Sources Identified

The canonical img2img + IPAdapter + ControlNet pipeline had three noise sources degrading quality:

The Winning Recipe: Poser-Only

sister_canon_poser.json — pure txt2img + DWPose ControlNet + hyper-detailed prompt:

Txt2img (clean latent)
  → DWPose skeleton from reference (ControlNet @ strength=1.0)
  → dpmpp_2m + karras @ CFG 5.5, 30 steps, denoise=1.0
  → Hyper-detailed prompt (every outfit detail, pose, expression)
  → No IPAdapter, no img2img, no face/outfit refs

Model Comparison

Pose Success Matrix

Saved Workflows

Workflow Architecture Comparison

Critical: Naming Confusion Risk

⚠️ sister_canon_pose.json (img2img + IPAdapter + DWPose — LEGACY) and
sister_canon_poser.json (txt2img + DWPose only — BREAKTHROUGH) have nearly identical
names but are architecturally opposite. Double-check filenames when selecting workflows.

Files Modified/Added This Session

Next Steps

• Lock poser recipe into --mode poser wrapper flag — Wire sister_canon_poser.json into sister_selfie.py. Add "poser" entry to MODE_DEFAULTS. Update prompt_engine.py to support mode="poser".
• Fix sister_experimental.sh — Either rename it (it wraps neutral-ref, not poser) or update it to use sister_canon_poser.json.
• Test Rei + Nei with poser recipe — Since workflows are generic (not per-sister), this is just sister_selfie.py --sister rei --mode poser
• Add poser mode to prompt_engine.py — Currently mode param only handles canon/selfie/stylized/glitch/nsfw. Poser mode should use QUALITY_BOOST_T2I instead of QUALITY_BOOST_IMG2IMG.
• Decide --mode natural fate — Current natural mode uses neutral-ref img2img (still has VAE noise). Should be deprecated or re-purposed as alias for --mode poser.
• Add wiki index reference — Current index doesn’t reference sister-photo-pipeline.md
• Build pose reference library — Downloadable pose images for DWPose input
• Tune animagine settings (LOW PRIORITY) — Only useful if NSFW Phase 4 starts. novaAnimeXL is the clear winner.
• Council sync — This document constitutes the council review.

Priority Order Recommendation

1. --mode poser flag (highest — unlocks the breakthrough for daily use)
2. Cross-sister validation (should work, confirm with tests)
3. Fix wiki/script mismatches (correctness)
4. Decide --mode natural deprecation (cleanup)
5. Add wiki index reference (discoverability)
6. Pose reference library (quality improvement)
7. Animagine tuning (defer to Phase 4)