S — Agent Pipeline Reference

The C Agents

S's autonomous travel journal — full agent architecture

Daily Run — End to End

What happens every day at 14:00 UTC

C-TravelAgent

Arc Planner

weeklyPlanner · Sun 10:00 UTC · deterministic · schedules arcs + home-active gaps → tripSchedule

↓ or manual override

Manual Trip (optional)

trips.html · city + dates → tripSchedule

↓ 3 days before start

C-Planner

Trip Planner

Sonnet · 13:00 UTC · generates itinerary JSON

↓ next morning

⏰

dailyBlog fires

Cloud Function · 14:00 UTC · every day

↓

C-Weather

Weather Fetch

Open-Meteo API · 3 retries · Firestore cache fallback · geocode + 7-day conditions

↓

C-Fashion

Stylist

Sonnet · designs N outfits · dedup against prior days · weather-aware · flux_phrase + narrative_phrase

↓ inline enrichments assembled for C-Storyteller

inline

Storyteller Inputs

assembled inline — no separate agents

• Sensory Palette — buildSensoryPalette() deterministic · weather feel, sounds, smells, texture
• Return Visit Memory — loadReturnMemory() · prior-visit fragments from travelerWorld
• Disruption — dayBrief.disruption field from C-Planner (Imperfection Engine · ~20% of mid-trip days)
• Research — optional C-Brief web_search output (--research flag)

↓

Narrative revision loop · max 2 revisions · threshold 8/10

C-Storyteller

Storyteller

Sonnet · 200–350 word diary · voice + place · consumes all inline inputs above

↔

C-Critic

Narrative QC

Opus · 2 retries · scores 1–10 per dimension · voice, specificity, structure · revise if overall <8

↓

C-Photographer

Photographer

Sonnet · extracts 3 + 1 spare scene · writes Flux prompts · weather-aware

↓

Image generation · 3 scenes + spare · fal.ai 3 retries · C-QC 3 retries · smart retry swaps spare on persistent failure

Flux Ultra

Image Generation

fal-ai/flux-pro/v1.1-ultra · img2img · strength 0.72 · 3:4 · 3 retries

↔

C-QC

Image QC

Sonnet · 3 retries · vision · hair color lock, solo, anatomy, style · retry at 0.65 · spare scene swap

↓

C-BRoll

B-Roll Scenes

Sonnet · atmospheric establishing shots · text-to-image · 16:9 · no character · 3 retries

↓

C-Editor

Coherence Gate

Sonnet · 3 retries · multi-modal · narrative vs images · hard conflicts trigger rewrite

↓

C-BlogEditor

Layout Editor

Haiku · reads narrative + scene descriptions · places all images contextually · fallback: positional (40%/75%)

↓

C-Archivist

Archivist

firebase-admin · saves day to Firestore · blog.html reads from here

↓

C-Report

Daily Email

nodemailer · HTML digest → epochblognet@gmail.com

↓ last day only

Trip Archive

Haiku · archiveTrip() · tripSummary + openThreads + characterNote → travelerWorld

Mixed models: Sonnet for reasoning · Opus for critique · Haiku for fast tasks · Director wraps the loop when --director flag is set

Orchestration

Director

Narrative

Fashion

Vision / Image

Quality Control

Archive / Memory

External Model

Phase 0 — Trip Input

C-Planner

Trip Planner

Watches tripSchedule Firestore collection for upcoming trips. 3 days before departure, generates a full multi-day itinerary JSON — areas, anchor hotels, day arcs, outfit seeds. Saves to trips/{tripId}.json and kicks off the blog pipeline.

IN: tripSchedule → city, country, startDate, endDate | OUT: trip JSON → Firestore + local file

claude-sonnet-4-6
Scheduled Cloud Function
runs daily @ 13:00 UTC

Live

Phase 1 — Director (Agentic Loop)

Director

Orchestrator

Agentic orchestrator with tool-use loop. Autonomously decides which sub-agents to call and in what order. Holds per-day working state. Tools: research_locations, design_outfits, write_story, extract_scenes, generate_images, edit_narrative (6 tools, director.js:38–110). Up to 10 reasoning loops per day with fallback to sequential steps on tool failure. Enabled via --director flag; the hardcoded linear pipeline remains as manual/fallback path.

IN: trip JSON, traveler config, S's World | OUT: fully published day (narrative + images)

claude-sonnet-4-6
director.js · runDirector()
--director flag (optional)

Live

Phase 2 — Narrative

C-Weather

Weather Feed

Fetches real daily conditions for every day of the active trip using Open-Meteo's geocoding and forecast/archive APIs — no API key required. 3 retries with backoff. Falls back to Firestore weatherCache if API fails. Returns a weatherByDay map with temperature range, WMO condition code, precipitation, and sensory prose. Injected into C-Fashion, C-Storyteller, C-Photographer, Sensory Palette, and the email report.

IN: city, country, startDate, endDate | OUT: weatherByDay{ [dateString]: { temp, condition, prose } }

Open-Meteo API
No LLM · free · no key
blog-weather.js

Live

Runs in parallel — Research & Fashion design

C-Brief

Travel Desk

Web search agent. Queries real information about the day's locations — architectural details, cultural context, local textures. Builds a researchContext block for C-Storyteller to use. Prevents generic hallucinated descriptions.

IN: location list | OUT: researchContext JSON

claude-haiku-4-5
web_search tool
--research flag

Optional

C-Fashion

Stylist

Designs S's outfit for each scene, grounded in her styleProfile DNA. Provides two phrases per scene: a Flux-ready visual description for image generation, and a narrative phrase for prose. Avoids outfit repetition by loading prior-day outfits from Firestore.

IN: styleProfile, dayBrief, priorOutfits | OUT: outfitBriefs[] per scene

claude-sonnet-4-6
styleProfile from
traveler-config.js

Live

Narrative revision loop max 2 revisions · threshold 8/10

C-Storyteller

Storyteller

Writes S's diary entry in her established voice — spare, perceptive, rooted in place. Uses research context for location authenticity and C-Fashion's outfit phrases to weave clothing into the prose naturally.

IN: dayBrief, outfitBriefs, researchContext | OUT: raw narrative draft

claude-sonnet-4-6

Live

C-Critic

Narrative QC

Scores the narrative 1–10 on each dimension (voice fidelity, location specificity, structure, emotional authenticity). Overall score 1–10, threshold 8. Returns summary, score, suggested revisions, and critique type (content/voice/structure). C-Storyteller rewrites; loop repeats up to 2 revisions.

IN: narrative + scoring rubric | OUT: critiqueSummary, critiqueScore, suggestedRevisions, critiqueType

claude-opus-4-6
blog-prompt.js:369
2 retries

Live

Phase 3 — Scene Extraction

C-Photographer

Photographer

Reads the finalized narrative and extracts 3 + 1 spare scene briefs — each with a location description (architectural, never just a name), shot type, lighting, secondary elements, and a Flux-ready image prompt with S's outfit embedded. The 4th spare scene is held in reserve for smart retry on QC failures.

IN: narrative, outfitBriefs | OUT: scenes[] JSON

claude-sonnet-4-6
blog-prompt.js:636
upgraded Phase 1

Live

Phase 4 — Image Generation

Flux Ultra

Single-Stage

Default image path. Takes C-Photographer's scene prompt + S's canon reference image. Runs img2img at strength 0.72 (locked) with 3 retries. Blank detection (<15KB) triggers retry at 0.65 fallback. Smart retry swaps spare scene on persistent QC failure. Uploaded to Firebase Storage for permanent URLs.

IN: scene prompt, canonUrl, strength | OUT: image URL (fal.media CDN)

fal-ai/flux-pro/v1.1-ultra
img2img · strength 0.72
aspect 3:4 · retry at 0.65

Live

Phase 5 — Quality Control & Coherence

Runs in sequence — Vision QC then narrative coherence check

C-QC

Image QC

Vision model inspects each generated image. Checks: solo subject (no co-subjects), S present, no anatomical errors, style match. Verdict: pass / retry / flag. On retry, regenerates the scene up to 3 times before flagging.

IN: image URL, scene brief | OUT: verdict + reason

claude-sonnet-4-6
vision · url source
blog-qc.js:34
3 retries w/ backoff

Live

C-Editor

Coherence Gate

Multi-modal editorial check. Loads narrative + all published images. Flags: existence conflicts (narrative denies photography), outfit mismatches, weather contradictions, mood mismatches. Hard conflicts trigger narrative rewrite; soft conflicts may flag scenes for regeneration.

IN: narrative + image URLs | OUT: conflicts[], revisedNarrative, scenesToRegenerate

claude-sonnet-4-6
multi-modal (vision)
--skip-editor to bypass
3 retries with backoff

Live

C-BlogEditor

Layout Editor

Reads the final narrative and all scene/B-roll descriptions. Decides contextually where each image should appear in the text — matching scene moments to the paragraphs that describe them. Character images and B-roll are woven into the reading flow. Falls back to mechanical positional placement (hero + 40% + 75%) if the agent fails or returns no data.

IN: narrative + scene descriptions + broll descriptions | OUT: placements[], narrativeFlow

claude-haiku-4-5
text-only (no vision)
blog-layout.js
2 retries · fallback: positional

Live

Phase 6 — Archive & Memory

Archive runs in parallel

C-Archivist

Archivist

Saves completed day data to Firestore travelerBlog collection. Stores narrative, scenes, image URLs, QC verdicts, outfit briefs, critic scores, and generation metadata. The source of truth for the blog reader.

IN: complete day object | OUT: Firestore write

firebase-admin
Firestore write
blog-firestore.js

Live

C-Memory

World Memory

Updates S's World — the persistent cross-trip memory in Firestore world collection. Tracks places visited, people encountered, recurring motifs, emotional threads. Referenced by C-Storyteller for continuity across trips.

IN: day summary | OUT: world Firestore update

claude-haiku-4-5
world-firestore.js
cross-trip memory

Live

C-Report

Daily Email

Fires at the end of every dailyBlog run. Builds a dark-theme HTML digest with pipeline status, S's hero image, real weather, outfit briefs, agent verdicts, narrative excerpt, and runtime metrics. Sent via Nodemailer to AB's project inbox.

IN: trip, dayDoc, weather, pipelineStart | OUT: HTML email → epochblognet@gmail.com

nodemailer + Gmail
GMAIL_USER secret
blog-report.js

Live

Autonomous Travel Agency

Zero-input trip planning — C-TravelAgent picks destinations, schedules arcs, manages home gaps

⏰

weeklyPlanner fires

Cloud Function · Sundays 10:00 UTC · checks if next arc is needed

↓

C-TravelAgent

Travel Agency

105 cities · seasonal scoring · novelty + recency + region variety · transit time from NYC · arc companions

↓ schedules 2–3 city arc

Firestore

tripSchedule

Writes city + dates → C-Planner picks up 3 days before → fully autonomous

↓ between arcs

⏰

dailyHome fires

Cloud Function · 14:00 UTC daily · checks for home-active periods · Mon/Wed/Fri only

↓

Home Mode

NYC Home Entries

150–250 words · 5 entry structures · real NYC weather · 1 B-Roll image · no character

Three-state rhythm: arc (full travel) → home-active (NYC Mon/Wed/Fri) → home-rest (dark)

Agent Index

Quick reference — all agents, models, and entry points

Agent	Role	Model	Entry Point	Triggered by
C-Planner	Trip Planner	claude-sonnet-4-6	blog-c0.js	Cloud Function scheduler · tripSchedule Firestore
Director	Orchestrator	claude-sonnet-4-6	director.js · runDirector()	--director flag in run-blog.js
C-Storyteller	Storyteller	claude-sonnet-4-6	blog-prompt.js · buildDayStory()	Director tool: write_story / doStory()
C-Critic	Narrative QC	claude-opus-4-6	blog-prompt.js · critiqueNarrative()	Inside C-Storyteller loop · 2 retries
C-Photographer	Photographer	claude-sonnet-4-6	blog-prompt.js · extractScenes()	Director tool: extract_scenes / doScenes()
C-Brief	Travel Desk	claude-haiku-4-5	blog-prompt.js · researchLocations()	Director tool: research_locations / --research flag
C4 (Flux)	Image Gen	fal-ai/flux-pro/v1.1-ultra	blog-generate.js	Director tool: generate_images / doImages()
C-Weather	Weather Feed	Open-Meteo API	blog-weather.js · fetchTripWeather()	Start of dailyBlog, before C-Fashion
C-QC	Image QC	claude-sonnet-4-6	blog-qc.js · checkBlogImage()	After each image generation · 3 retries
C-Fashion	Stylist	claude-sonnet-4-6	blog-fashion.js · designDayOutfits()	Director tool: design_outfits / doFashion()
C-Report	Daily Email	nodemailer	blog-report.js · sendDailyReport()	End of dailyBlog, after C-Archivist save
C-Editor	Coherence Gate	claude-sonnet-4-6	blog-editor.js · runEditPass()	Director tool: edit_narrative / doEdit() · 3 retries
C-BlogEditor	Layout Editor	claude-haiku-4-5	blog-layout.js · runLayoutPass()	After C-Editor · --mode=layout for manual · fallback: positional
C-Archivist	Archivist	firebase-admin	blog-firestore.js · saveDay()	End of each day loop, auto
C-Memory	World Memory	claude-haiku-4-5	blog-prompt.js · archiveTrip()	Last day of trip only, after saveDay()
C-BRoll	B-Roll Scenes	claude-sonnet-4-6	blog-prompt.js · extractBrollScenes()	After character images · text-to-image 16:9
C-TravelAgent	Travel Agency	Deterministic (no LLM)	travel-agent.js · planNext()	weeklyPlanner Cloud Function · Sundays 10:00 UTC
Home Mode	NYC Entries	claude-sonnet-4-6	blog-home.js · runHomeEntry()	dailyHome Cloud Function · Mon/Wed/Fri during gaps

Firestore Collections

Data architecture — what lives where

tripSchedule

{tripId}

city · country
startDate · endDate
status (planned/running/done)
→ Written by: trips.html + C-TravelAgent
→ Read by: C-Planner

travelerBlog

{tripId} / days / {dayN}

narrative · scenes[] · spareScene
images[] · brollImages[]
outfitBriefs · criticScore
narrativeCriticError · disruption
→ Written by: C-Archivist
→ Read by: blog.html, dashboard.html

travelerRoom

{travelerId}

canonUrl · styleProfile
name · origin · voice
→ Written by: traveler.html
→ Read by: C-Storyteller, C-Fashion

travelerWorld

{travelerId}

visitedCities[] · tripHistory[]
openThreads[] · characterNote
returnVisitFragments
→ Written by: C-Memory
→ Read by: Return Visit Memory

weatherCache

{city-slug-startDate}

weatherByDay{}
cachedAt timestamp
→ Written by: blog-weather.js
→ Read by: fallback on API fail

travelerBlog/nyc-home/days

{YYYY-MM-DD}

narrative · brollImages[]
entryType · weatherContext
tripSchedule doc: nyc-home-{startDate}
→ Written by: blog-home.js
→ Read by: blog.html

CLI Run Modes

node src/run-blog.js — common invocations

Command	What it does
--mode=trip --trip=osaka-2026-03 --director	Full agentic run — Director orchestrates all agents autonomously (primary mode)
--mode=trip --trip=osaka-2026-03	Linear fallback pipeline: fashion → story → scenes → images → QC → editor → layout → save
--mode=trip --trip=osaka-2026-03 --day=3	Run a single day of the linear pipeline end-to-end
--mode=plan --trip=osaka-2026-03	Run C-Planner (C0) to generate the trip JSON
--mode=brief --trip=osaka-2026-03 --day=3	Run C-Brief Travel Desk web_search research for a day's locations
--mode=fashion --trip=osaka-2026-03 --day=3	Run C-Fashion only — preview outfit briefs
--mode=story --trip=osaka-2026-03 --day=3	Run C-Storyteller + C-Critic loop only
--mode=scenes --trip=osaka-2026-03 --day=3 --scenes=3	Run C-Photographer scene extraction (default scenes=3)
--mode=generate --trip=osaka-2026-03 --day=3	Generate images from existing scenes (Flux Ultra img2img)
--mode=broll --trip=osaka-2026-03 --day=3	Extract + generate B-roll environmental scenes (no character)
--mode=qc --trip=osaka-2026-03 --day=3 --force	Re-run C-QC on existing images; --force re-checks prior passes
--mode=edit --trip=osaka-2026-03 --day=3	Run C-Editor coherence gate only on saved day
--mode=layout --trip=osaka-2026-03 --day=3	Run C-BlogEditor — compute smart image placement
--mode=archive --trip=osaka-2026-03	Run C5 / archiveTrip on a completed trip (tripSummary + openThreads)
--research	Enable C-Brief web_search during full runs (feeds C-Storyteller)
--skip-editor	Bypass C-Editor coherence gate (speed mode)
--no-weather	Skip C-Weather fetch (dev/offline runs)
--two-stage	Two-stage image pipeline: gpt-image-1 → flux-fill inpaint/img2img
--dry-run	No Firestore writes, no image generation (smoke test)