Session-scale dependency map (mandatory audit, spec #2/#3)

Source: full consumer audit of DSH HEAD (0.1.1-rc.2, b150a551b8). Key: FULL = needs whole array · SUFFIX = range/slice · RA = random access by seq · PROJ = fold/projection · IDENTITY = exact array/object identity matters · SYNC = breaks if made async.

The decisive finding

deriveMessages() never reads raw deltas. packages/core/session/src/index.ts:729 walks only surface nodes incrementally (surface.nodes, cursor-based, per-call cost = new nodes only; rebuild on rewrite). Projection (core/session/src/surface.ts:83–114) consumes exactly: user/message, non-empty assistant/message, tool/result. Chunk/boundary/log-only events derive to null.

Consequence for the giant fixture: ~1.09 M packed chunk events are pure materialization overhead for agent-ready resume. They matter for replay/UI fidelity, not for continuing execution.

Related incremental folds already exist inside Session: requestHeader() (index.ts:673–683) and requestContext() (694–702) — same cursor pattern.

Consumer classification

Consumer	Path	Access	Notes
Session _forkSeed	core/session index.ts:1101–1140	RA + SUFFIX(prefix) + findLast	seed boundary semantics
Agent Loop	core/agent-loop agent.ts:92 findLast('turn/start'); runtime-context.ts:36–44 reverse scan w/ surface.nodes	recent-event scans SYNC	dev-only invariant does FULL fold
agent inbox	core/agent/src/inbox.ts:32	SUFFIX from seedLength, one-time iterate
Token meter	llm/token-meter index.ts:174–179,297	incremental ITER + RA indexed by seq	portable to indexed source
llm-retry	llm-retry index.ts:182; invariant.ts:150–164	recent RA + FULL(dev-only)
Goal	goal index.ts:425,438; tool-goal authority.ts:31	PROJ incremental + suffix	full seed fold once
Schedule	schedule index.ts:51; tools.ts:224; runtime.ts:209	FULL arg	caching unverified (unknown)
Wait folds	wait/wait index.ts:358	FULL arg → single fold	durable wait history independent
Teams (experimental)	experimental/agent-team journal.ts:31; mailbox/roster slice(seedLength)	FULL fold + SUFFIX
Plan mode	plan-mode index.ts:249…501	FULL re-fold PER CALL sync	hot, folding candidate
Compaction	compaction-basic region.ts:163,306,503–507; tool-pairing.ts:59	RA-by-seq + reverse scans	keying by shadowedSeqs
Coordinator itself	coordinator.ts:742,918,975,984,989,1174,1196	length check / snapshot borrow / suffix clone	owns current guarantees
session-projection	session/session-projection index.ts:457–494	PROJ with checkpoint restore + FULL fallback	ALREADY has checkpoint persistence! (444–446)
Session title	session-title index.ts:350…760	FULL folds + findLast
Telemetry	session-telemetry coordinator.ts:142 adoption replay; otel index.ts:247	FULL iter + IDENTITY: compares event refs to session.events[seq]	identity requirement lives here specifically
Session Query	session-query corpus.ts:242,295; session-query-sqlite index.ts:856–870,790–835	FULL clone per observation + FTS union	heaviest quadratic-ish consumer on giant sessions
ApiProxy/host	apiproxy api-proxy.ts:451,498–504,1384–1404,3494	some() / list-metadata fold / recent pairing scan; export uses RAW artifacts not .events (:3580+)
Subagent drivers	subagent*/lifecycle.ts:196 etc	boundary=length; own suffix slices

Cross-cutting contract facts

• session.events is a cached frozen copy of the internal log until next append (core index.ts:551–562); events deep-frozen at acceptance (append :630, restore freezeRestoredObject :536).
• Publish path: coordinator prepareCore → ctx.sessions.prepare(id,{seed,…}) (coordinator.ts:905); Session.fromRestore validates envelopes then freezes.
• Prepared-session cache: count-based, default size 5 (coordinator.ts:27, preparations.ts LRU). No byte-weight budget exists (#36 gap).
• Everything listed is synchronous today; the request path calls deriveMessages() inline (agent-loop agent.ts:341).
• Unknowns flagged honestly: schedule fold caching; whether client-runtime hydrates old trajectory turns from persistence on attach.

Implications for paged hydration

1. Agent resume needs NO historical chunks — a surface-node/descriptor-grade representation suffices (spec #27/#28 confirmed structurally, not just hoped).
2. Sequenced consumers (token meter, telemetry otel identity check) are the ones that pin array-indexed identity — they must either keep compatibility arrays or migrate to seq-coordinate access explicitly (#91 vs #90 advice honored).
3. session-projection ALREADY implements its own durable checkpoint + FULL fallback (index.ts:444–446) — precedent showing a checkpointable-fold pattern exists in core vocabulary (spec #67).
4. Session Query is the best first independent optimization target after any seam lands: it re-clones whole histories per observation today (#93/#48).