Scratchpads and handoffs

A scratchpad is where discovery work can stay provisional. It is not a smaller canon database and not a disposable chat transcript. NovelGraph stores it as a time-ordered record tied to one discovery session, so hypotheses, conflicts, alternatives, and questions can inform the next handoff without silently becoming a decision.

The canonical agent-boundaries diagram describes the route from author input through Sol, Terra, and Luna to a context dossier and author decision.

Textual equivalent:

Author → Sol coordinator → Terra deep analysis or Luna bounded work
                           ↓                         ↓
                     session scratchpad ←─────────────┘
                           ↓
                   role-specific dossier → author approval or revision

What belongs in a scratchpad

Every scratchpad entry has an agent role, kind, content, optional confidence, source references, and timestamp. Its permitted kinds are deliberately modest:

Kind	Appropriate use
observation	A compact statement noticed in a turn or source
hypothesis	A possible explanation, pressure, or story mechanism
question	A decision that needs author input or further evidence
contradiction	Two claims or directions that cannot both remain true as stated
alternative	A distinct route worth comparing, not merely a restatement
digest	A traceable summary of earlier exploratory material

Source references matter even for internal material. An entry can point back to the turn, claim, prior entry, or other identifier that motivated it. Confidence is optional because an agent should be able to say “this is a question worth asking” without inventing numerical precision. When confidence is used, the input accepts values from 0 through 1.

The store keeps scratchpad entries in creation order. It does not deduplicate them, automatically promote them, or erase them when a later entry disagrees. A good handoff can refer to a contradiction and explain its resolution; it should not overwrite the fact that the contradiction existed.

Roles create different kinds of handoff

Sol is the coordinator for the author’s conversation. It can read discovery, scratchpad, literary, series, and book material; it can write discovery records, propose charters and canon changes, and request approval. Its job is to choose the next consequential question and return decisions to the author, not to approve them.

Terra is the specialist route for structural judgment. It can read the same discovery, literary, series, and book context and can add scratchpad material or propose a charter. It does not receive the canonical-promotion or approval-request capability. That separation makes a useful review pattern: Terra can expose a weak character engine, a genre obligation, or an ending risk; Sol can present the resulting proposal; the author decides whether it governs the book.

Luna is a bounded worker. It can read discovery, scratchpad, literary, and book context, then write scratchpad entries. Its capability set excludes series reading and charter proposals. That is suitable for tasks such as extracting named entities from a turn, locating an explicit constraint, comparing two supplied alternatives, or identifying an unresolved observation. A smaller assignment should not arrive with unrelated continuity or authority.

Handoff	Include	Do not imply
Luna to scratchpad	Extracted observations, source references, uncertainty	That an extraction is a book fact
Terra to Sol	Structural analysis, alternatives, risks, charter-ready language	That Terra can approve or promote it
Sol to author	The decision, supporting context, and consequences	That a pending approval is already canon
Author decision to system	Approval or rejection with rationale	That rejected material disappears from history

What a dossier actually contains

DiscoveryEngine.dossier(bookId, agentRole) assembles a role-labeled snapshot and stores it in context_dossiers. Its content includes the latest approved charter, the knowledge bases available to the book, all approved claims from those bases, the current session’s scratchpad entries, unresolved discovery observations, and the capabilities for the requested role.

The important qualification is “approved claims.” A dossier does not promote a working note just because it is present in the session. At the same time, it includes unresolved observations so an agent can identify the question that needs attention rather than treating the approved charter as an exhaustive account of the book.

The current implementation gathers approved claims from the bases owned by the book or linked to it. It writes a new dossier record whenever a session exists. Dossiers are snapshots, not a live view that can be silently changed after a handoff. If a proposal later needs audit, the stored content shows what information and permissions framed it at that point.

Turning exploration into a decision

The canonical canon-promotion diagram shows the intended route from conversation through extraction and synthesis to a typed proposal, author review, and either approved canon or a rejected direction.

Textual equivalent:

Conversation → extraction → scratchpad → synthesis → typed proposal
                                                        │
                                      author approves ──┼── author rejects
                                                        ↓
                                                   approved claim

There are two formal decision routes in the discovery core. A proposed Story Charter creates a pending approval and requires a rationale on resolution. A proposed knowledge claim is promoted through promoteClaim, which also requires a rationale and records an approval identifier. Neither route turns scratchpad content into canon automatically.

That last rule matters most when an idea is repeated. Repetition may make a hypothesis more interesting; it does not change its provenance or status. A digest can make a long investigation readable, but it should link back to the entries it summarizes and preserve rejected alternatives when they explain why the chosen route was preferred.

Practical handoff discipline

Before giving a task to another role, state the decision it is meant to inform, the sources it may use, and the status of each important statement. Return a result that separates facts, inferences, options, and questions. If a new premise is proposed, state what it would constrain or foreclose. If a contradiction is found, name both sides rather than quietly selecting one.

This produces a trail that remains useful after a draft changes direction. The scratchpad is where a book can think in public to its own project history. Canon is where it commits.