Query, not chat: when managed context appends and when it reseeds

A managed fak session is one continuing query against a shared task model, not an append-only conversation. The query may run for many turns and workers, but every turn must still sharpen the same task or reconstruct it from durable evidence. This page is doctrine: it states the rule that gateway and resume work should implement; it does not claim that every reseed decision is already enforced automatically.

1. Doctrine: one query, not an ever-growing chat

Chat treats the transcript as the product: each new message inherits every earlier message simply because it came later. Managed context treats the task as the product. History is useful only while it improves the current task model.

The operational question for every turn is therefore not “can this message be appended?” but “does this information sharpen the query we are still answering?” Append when it does. Reseed when a clean reconstruction from durable state is more faithful or cheaper than carrying the conversational residue forward.

2. Pin the originating task

Each managed session has one originating-task pin: a stable, content-addressable statement of the task whose outcome the session is pursuing. The pin is the identity of the query, not merely the oldest user message and not a summary of the latest failure. Workers may refine an execution plan, but they must be able to trace that plan back to the same pin.

A turn that changes or replaces the originating task starts a new query and therefore requires a reseed. Compaction and restart must preserve or restore the pin before work continues. If the pin cannot be recovered, the session is not entitled to infer it from stale conversational fragments.

3. Keep stable state; swap transient state

Stable state survives append, compaction, reseed, and restart:

• the originating-task pin;
• accepted constraints and decisions that still govern the task;
• durable witnesses, checkpoints, and externally readable effects; and
• the next action when it is justified by those witnesses.

Swappable state is useful only while it changes the next decision:

• the current error, tool result, or provider residue;
• transient hypotheses and failed approaches;
• superseded plans; and
• narration whose effect already exists as a durable witness.

Swappable does not mean unimportant. It means replaceable without changing the query’s identity. A reseed reconstructs the current query from stable state and includes only the swappable state still needed for the next action.

4. Append only task-relevant witness

Append only when the new turn adds task-relevant witness that sharpens the same originating-task pin. A valid append must answer all of these questions positively:

• Is the originating task unchanged?
• Does the new information alter or justify the next action?
• Is it still current rather than superseded by a later witness?
• Is retaining it cheaper and clearer than reconstructing the current state?
• Can another worker explain why this turn belongs to this task without relying on chat chronology alone?

Examples include a fresh test result, an adjudicated decision, a newly observed external effect, or a constraint that narrows the same task. Repetition, status narration, and an error already captured in a checkpoint are not reasons to append.

5. Reseed from durable positive state

Reseed when conversational history no longer sharpens the pin. Any one of these conditions is sufficient:

• the turn changes or replaces the originating task;
• stale errors, results, or abandoned approaches dominate the useful state;
• context pressure makes retained history costlier than reconstructing from durable state;
• compaction, restart, or amnesia threatens the originating-task pin; or
• the next action can be reconstructed from the task plus durable witnesses without the conversational residue.

Reseed means constructing a new, positive current-state query: originating task, applicable constraints, confirmed effects, current decision, and next checkable action. It does not mean merely deleting old turns, summarizing their chronology, or adding a new instruction that says to ignore them.

6. Checkable reseed-versus-append procedure

At each managed-context boundary, evaluate this list in order:

1. Read and identify the originating-task pin. If it is missing or changed, reseed.
2. Read durable witnesses and checkpoints independently of the transcript.
3. Remove superseded errors, results, hypotheses, plans, and narration from the candidate state.
4. Ask whether the proposed turn adds current witness that changes or justifies the next action for the same pin.
5. If yes, and retaining it costs less than reconstruction, append it.
6. Otherwise, reconstruct task + constraints + confirmed effects + next action and reseed.
7. Verify that a worker receiving only the reconstructed query can name the same pin and take the same justified next action. If not, the reseed is incomplete.

This procedure yields an observable decision: the boundary can name the pin, the durable witnesses retained, the residue dropped, and the reason for APPEND or RESEED.

7. Shared workspace, positive state, and existing seams

The managed query is a global-workspace surface: workers read and update one shared model of the task rather than maintaining independent chat stories. Durable effects and checkpoints are the shared facts; a worker’s local transcript is only a temporary view. This lets a replacement worker resume from the same pin without inheriting another worker’s stale reasoning.

The positive-state rule matters because model-visible text is broadcast state. Repeated negations such as “ignore the old error” keep the old error resident and require the model to invert it again on every turn. Reseeding asserts what is true now and removes the obsolete negative operand. Appending remains valuable when it contributes new evidence, but append-only drift accumulates broadcast contamination and negation tax.

Existing fak seams provide parts of this contract:

• internal/gateway/ctxrestore.go restores a compacted originating task by content-addressed identity;
• internal/agent/anthropic_compact.go records and restores the origin-task tombstone across provider compaction;
• internal/sessionsteer/sessionsteer.go exposes checkpoint and rebuild directives;
• internal/gateway/ctxvalue.go carries managed-context step advice; and
• internal/negframe supplies the positive-state reframe boundary.

These are implementation footholds, not a claim that automatic reseed-versus-append enforcement is complete. Future enforcement should make the procedure above mechanical while preserving this doctrine’s one invariant: every surviving turn must sharpen the same originating-task pin.