Story Computing Machine

A small shared-universe story room. A [WP] prompt becomes four authored files and, after review, one title image. An explicitly named replacement removes the old package and creates the same clean shape while Git preserves its history:

prompt.md → outline.md → story.md → review.md → title-image.jpg

The coordinator delegates fresh outliner, writer, and reviewer agents for each story. The writer uses one local production adapter for in-place structural, dialogue, and language revision; that writer may also make a targeted REVISE pass, while every later review is fresh. The reviewer reads the prose before the outline and checks people, places, prompt fulfillment, dialogue, binding narrative policy, and continuity. For new stories it also checks recent-story interchangeability. Clean CREATE prompts derive a dialogue promise, dialogic medium, dialogue engine, and relationship movement before drafting; the illustrator reads the finished prose and creates an exact 864x1536 portrait novel cover. Reference images attached to the prompt are inspected and supplied to both the outliner and the cover generator; only their display names are recorded in prompt.md, and the source images remain outside the story directory. There are no draft/final duplicates, canon briefs, handoff ledgers, or release certificates. Each cover includes the exact reader-facing story title once.

Quick start

1. Open the primary main checkout as the Codex workspace.

2. Submit a prompt such as:

   [WP] Every city has a ghost assigned to it. Tonight, ours resigns.
   Target: about 3,000 words; close third person; melancholy but hopeful.
   

   Attach any character, setting, object, mood, palette, or style references to that same request. Every supplied image will inform the outline and be passed as a reference input when generating the cover.

3. The coordinator updates main, creates codex/story-<slug>, adds a dedicated sibling Git worktree for the branch, and runs the complete story workflow from that worktree.

The default result is a reviewed, non-canon story. Adding it to shared canon is a separate explicit user decision.

Repository map

• AGENTS.md — the complete operating rules.
• .agents/skills/story-create/ — the user-facing new-story and replacement coordinator.
• .agents/skills/story-room/ — shared outline, review, and title-image stage contracts plus scaffold and validation scripts.
• .agents/skills/short-story-writing/ — the compact project-owned production adapter that resolves upstream craft advice into one in-place prose workflow.
• .agents/skills/creative-writing-craft/, dialogue/, and prose-style/ — pinned upstream craft references; the first is primary and the other two are diagnostic lenses inside the local adapter.
• .agents/skills/story-analysis/, story-sense/, and sensitivity-check/ — pinned upstream review references used within the repository review contract.
• skills-lock.json — upstream skill sources, immutable commit refs, and content hashes for restoration and updates.
• .codex/agents/ — the narrow outliner, writer, reviewer, and title-image roles.
• stories/_template/ — the four-file scaffold.
• universe/ — authoritative shared-universe facts and style constraints.
• stories/ — bundle-format and current-format story packages; title images live beside their story prose as title-image.jpg.
• pages/catalog.json — the stored publication snapshot used by Pages.
• pages/covers/ — captured title images used by the index and story pages.
• pages/timeline.json — the curated in-universe era map and placement confidence used by the visual chronology.
• pages/build.py — captures reviewed stories and renders the snapshot.

Two validation phases

pwsh -NoProfile -File .agents/skills/story-room/scripts/Test-Stories.ps1 -Story <slug> -Phase PreReview
pwsh -NoProfile -File .agents/skills/story-room/scripts/Test-Stories.ps1 -Phase Final

The agent runs the targeted pre-review phase before review, generates the title image only after a passing review, and then runs the final phase locally. Final validation requires a readable 864x1536 JPEG for every current story. For prospective-2026-08-23, local validation also enforces the compact outline ceiling, the three dialogue-design fields, the six-field Voice capsule, and the existing structured dialogue verdict. The earlier 08-18 and 08-21 contracts remain valid unchanged; 08-21 retains its five-field, 180-word Voice capsule. Both phases validate current-format prose. Final validation also inventories bundle-format packages and verifies that each has a boolean canon flag; semantic noun extraction, dialogue judgment, and continuity judgment remain the reviewer's responsibility.

Replacing a completed story

A whole-story remake, rewrite, overwrite, or replacement uses the same story-create workflow as a new story. It is available only for an explicitly named package. The coordinator creates codex/story-<slug> and a dedicated worktree before inspecting the target's canon marker or story inputs.

A canon story must first receive a separate marker-only unlock patch and commit. For an editable target, the coordinator preserves every verbatim user-authored prompt or request block, the new request, and every associated reference-image display-name inventory. Machine selection, cover-policy, and workflow metadata are discarded. The images must still be available as external paths or attachments; a display name cannot recover an image, so inaccessible originals must be reattached.

The old package and its catalog entry, captured cover, chronology placement, and bundle index row are then removed. stories/NAMES.md remains frozen production memory, and authoritative universe facts are never removed as part of a replacement. Once the target path is absent, new-story.ps1 creates a clean four-file package with the preserved prompt text and all accessible references. The outliner, writer, and reviewer receive no prior outline, prose, review, or cover. A replacement receives the ordinary CREATE craft profile, recent-story comparison, fresh cover, final validation, capture, parity check, push, and draft pull request.

A narrowly requested edit to one named non-canon bundle-format story may still modify that bundle's prose directly under AGENTS.md's canon-state rules. Completed packages with historical rewrite sections remain readable, but those sections are inert history and are never produced by the current workflow.

Pages

Pages does not read stories/ during its build. Every passing story is handed to the stored catalog once:

python pages/build.py capture <slug>

Capture requires a passing review but does not duplicate the full story check. It stores the prose in pages/catalog.json and copies the title image into pages/covers/. Capture and capture-all refuse to demote a catalog-canon story when its source marker disagrees; reconcile that marker through the named canon process first. The Pages index renders cover cards with each story's title, prompt, cover, created and edited dates, state, word count, and content rating; each story page places the cover below its title and prompt. The separate timeline.html page arranges every cover by its proposed place in the universal chronology—not publication or reading order—from the first wonders through the Long Zero and Joined-Sky reawakening. Solid, dashed, dotted, and open-split rings distinguish fixed, inferred, speculative, and unresolved placement confidence. Numbered era nodes order the eras; story groups inside each chapter are explicitly unordered. The chronology is a reader-facing working model; universe/ remains the shared fact authority. It uses the same captured cover files as the library and story pages. The check command requires exact parity among source packages, the bundle index, catalog, captured-cover bytes, chronology placements, and authoritative source/catalog canon flags. Verify the snapshot and reader locally:

python pages/build.py check
python -m unittest discover -s pages -p "test_*.py" -v
python pages/build.py build --output _site

After merge, GitHub Actions only builds, uploads, and deploys the stored catalog, chronology map, captured covers, reader stylesheet, and chronology interaction script; it does not rerun story validation or reader tests.