Oh Story review Chapter checks and tracking consistency
A disposable Chinese-fiction project tests Oh Story length gates, explicit exceptions, state revisions and derived-file consistency without model generation.
AgentIndex · Published 2026-11-05 · Updated 2026-11-05
Oh Story packages Chinese web-fiction workflows as agent skills, covering research, outlines, prose, checks and publication preparation. For long-form authors, its most assessable contribution is how a completed chapter becomes durable progress that another session can recover from files.
Our chapter and tracking checks found useful boundaries: valid fixtures advanced state, short chapters required an explicit decision, quality failures and stale revisions stopped commits, and manual changes to derived files were detected. These properties help maintain project records. They do not establish that a novel is engaging or its characters believable.
Version and twelve local checks
We pinned commit 0b5cf2a7a37d605706f6406b66526f6d6d714141, whose plugin manifest reports 0.10.0. The MIT repository contains fourteen skills. Tests ran with Python 3.12.13 and Node.js 26.6.0, without installing skills or loading hooks into an everyday host. Project description
The input was a temporary novel project. Repeated Chinese characters provided precise mechanical length specimens; they were not generated fiction or a literary benchmark. Actual storyctl.py and tracking_commit.py commands produced the observed files and statuses. The reproduction fixture records source fingerprints and results.
| Check | Observed behavior |
|---|---|
| 1000-character fixture and target | ready; check did not change tracking state |
| Normal commit | Advanced durable chapter state |
| Derived views after commit | Agreed with canonical state |
| 800-character fixture against 1000 target | needs_decision |
| Ordinary short-chapter commit | LENGTH_OUT_OF_BAND; progress unchanged |
| Explicit current-length acceptance | Committed and recorded exception |
| Known prose-pattern fixture | QUALITY_BLOCKED; no progress advance |
| Stale state revision | Refused overwrite |
| Manual context-view edit | Drift detected; restored file passed |
The script counts related operations separately for twelve passing checks. Writer invocation, cover generation and publication were not included.
Checking does not complete a chapter
The first fixture returned ready with an actual length of 1000. Its tracking JSON remained byte-for-byte unchanged. Only the subsequent commit advanced the completed chapter. This distinction lets an author inspect a result without mistaking inspection for durable acceptance.
Likewise, the existence of prose does not establish synchronized tracking. Before ending a writing session, confirm that chapter text, canonical state and derived views describe the same point in the story. Otherwise a later session may inherit an outdated commitment or location. Chapter-control implementation
Try a complete check, commit and recovery sequence on a disposable copy before expanding the workflow. That makes failures easier to locate than beginning with a large manuscript import. We did not evaluate old-book import or cross-device synchronization.
Record the decision to accept a shorter chapter
Our second chapter had 800 characters against a target of 1000. It required a decision, and ordinary commit was refused while chapter one remained the latest completed chapter. Explicit accept-current-length then committed it and stored accepted_current_length.
That separates a consciously accepted deviation from an unnoticed one. The command does not authenticate an author, however: a program able to execute it can make the same call. Host workflows should preserve the author's decision rather than treating a command name as an authorization boundary.
Length should not become the creative objective. Padding prose to satisfy a gate can harm pacing, while shortening a chapter can omit necessary information. The tool can record a deviation and its resolution; the author still decides whether the chapter works.
Canonical state and derived files
A third-chapter transaction using an old revision was rejected without changing canonical state. The current revision succeeded. We then edited 追踪/上下文.md manually; the consistency check found a difference from _tracking-state.json. Restoring the original view made the check pass. Tracking implementation
Derived Markdown is therefore not an independently maintained source of truth. Update lasting state through the expected transaction path and regenerate views. A manually improved-looking paragraph can still be inconsistent with the canonical record.
Revisions help detect an older session overwriting newer progress. They do not merge two authors' creative decisions automatically. Collaborative use still needs ownership of commits, a way to compare differences and recoverable backups. We did not run a distributed concurrency test.
Quality gates check particular rules
Adding one known prose pattern caused a quality failure, and ordinary commit returned QUALITY_BLOCKED. Removing the specimen and using a current revision allowed progress. This establishes that the selected rule participates in the gate; it does not establish detection of every repetition, contradiction or stylistic weakness.
Our repeated-character length specimens could pass when they did not trigger the tested blocking rules. That alone demonstrates the distance between passing mechanical rules and readable prose. Zero findings should not be presented as proof of originality, detector avoidance or the removal of every sign of generated writing. Character motivation, revealed information, payoff and pacing still require editorial judgment.
Adoption decision
Try a small project copy if you already keep settings and outlines in files and want a visible chapter-completion process. Verify hook support and execution separately in the intended host; we invoked scripts directly. An MIT skill package also does not include host-model or image-service costs.
This review provides no literary-quality, platform-approval or commercial-outcome result. Its practical case is clearer state management for authors willing to maintain canonical records and interpret rule feedback. Project and host entry point
For cross-session handoffs, see the Crew shared-log review. For organizing scripts into shots and materials, continue with the Toonflow canvas review.