united-workforce/docs/plans/2026-05-09-cas-thread-storage.md

RFC: CAS-Based Thread Storage

Status: Draft Author: 小橘 🍊（NEKO Team） Date: 2026-05-09

Summary

Replace .data.jsonl with a fully CAS-based thread state chain. Threads become linked lists of immutable CAS nodes, indexed by a per-bundle threads.json.

Motivation

.data.jsonl is a flat append-only file with three different row formats (start, role step, end). This makes forking expensive (copy file), deduplication impossible (forked threads repeat shared history), and GC complex (must parse every row to find CAS refs).

Threads are inherently immutable append-only sequences — a natural fit for CAS hash chains, similar to git's commit DAG.

Design

Node Types

Two CAS node types, using the existing { type, payload, refs } CAS blob structure:

StartNode

Contains workflow-level parameters. No threadId (because the same StartNode can be shared across forks). Prompt is stored as a CAS blob and referenced via refs[0].

CAS blob:
{
  type: "start",
  payload: {
    name: "solve-issue",
    hash: "BUNDLE_HASH",
    maxRounds: 10,
    depth: 0
  },
  refs: [
    <prompt_hash>    // refs[0]: initial task prompt (CAS blob)
  ]
}

• No role, content, meta — this is not a step, it's workflow metadata
• Prompt is not inline — it lives in CAS and is referenced by hash

StateNode

One per role step (including __end__).

CAS blob:
{
  type: "state",
  payload: {
    role: "coder",
    meta: { ... },
    start: "<start_hash>",
    content: "<content_merkle_hash>",
    ancestors: ["<parent_hash>", "<grandparent_hash>", ...],
    compact: null,
    timestamp: 1234567890
  },
  refs: [<start_hash>, <content_hash>, <parent_hash>, ...]
}

Payload is the source of truth. Application code reads named fields from payload. refs[] is a GC index — automatically derived from payload by collecting all CAS hashes. GC only scans refs[] without understanding payload structure.

Payload fields:

Field	Type	Meaning
role	string	Role name, or "__end__" for completion
meta	object	Structured metadata extracted from agent output
start	string	StartNode hash
content	string	Content Merkle node hash (carries role artifact refs)
ancestors	string[]	[parent, grandparent, ...] — up to 11 entries (1 parent + 10 skip-list). Empty for first step after start. ancestors[0] is the direct parent.
compact	string | null	CAS hash of a compacted summary of all nodes before this one. When present, LLM context assembly can use this instead of walking the full chain.
timestamp	number	Unix timestamp in ms

Content Merkle Node

The content at refs[2] of each StateNode is itself a CAS Merkle node. This is where role artifact references live:

CAS blob:
{
  type: "content",
  payload: "<role output text>",
  refs: [
    <artifact_hash_1>,   // e.g. a commit, a file, a sub-result
    <artifact_hash_2>,
    ...
  ]
}

The Extractor is responsible for producing both meta and refs from raw agent output:

Agent raw output
    ↓
Extractor → { meta, contentPayload, refs[] }
    ↓
CAS put content Merkle: { type: "content", payload: contentPayload, refs }
    ↓ contentHash
StateNode: { ..., refs: [start, parent, contentHash, ...ancestors] }

This keeps StateNode refs fixed and simple. All role-specific artifact references are encapsulated in the content Merkle node. GC follows: thread head → StateNode.refs → content Merkle.refs → artifacts, full chain recursive.

End Node

An end is just a StateNode with role: "__end__":

{
  type: "state",
  payload: {
    role: "__end__",
    meta: { returnCode: 0, summary: "completed successfully" },
    start: "<start_hash>",
    content: "<content_hash>",
    ancestors: ["<parent_hash>", ...],
    compact: null,
    timestamp: 1234567891
  },
  refs: [<start_hash>, <content_hash>, <parent_hash>, ...]
}

Thread Index: threads.json

Per-bundle directory, one threads.json file. Only active (in-progress) threads live here:

~/.uncaged/workflow/bundles/<hash>/threads.json

{
  "01JTHREAD1AAAAAAAAAAAAAAA": {
    "head": "<latest_state_node_hash>",
    "start": "<start_node_hash>",
    "updatedAt": 1234567891
  }
}

When a thread completes (__end__), it is removed from threads.json and appended to a date-partitioned history file:

~/.uncaged/workflow/bundles/<hash>/history/{YYYY-MM-DD}.jsonl

Each line:

{"threadId":"01JTHREAD1AAAAAAAAAAAAAAA","head":"<end_node_hash>","start":"<start_node_hash>","completedAt":1234567891}

Benefits:

• threads.json stays small — only in-flight threads
• Dashboard watches threads.json for real-time updates; completed threads don't trigger watches
• History is queryable by date but not actively monitored
• GC roots = all heads from threads.json + all heads from history/*.jsonl

Ancestor Skip-List

Each StateNode carries up to 11 entries in payload.ancestors (1 parent + 10 skip-list, newest first):

Node 15: ancestors = [node14, node13, node12, node11, node10, node9, node8, node7, node6, node5, node4]
                      ^parent  ^--- skip-list (10 most recent) ---^

This enables:

• Paginated fetch: jump to any recent ancestor without walking the full chain
• Partial replay: fetch last N steps without loading the entire history
• The list is capped at 10 to keep node size bounded

Fork

Forking a thread at step N:

1. Create new threadId
2. Create a new StateNode whose parent (refs[1]) points to the fork point's StateNode
3. Register the new threadId in threads.json with its own head
4. Zero data duplication — the forked thread shares all ancestor nodes via CAS

Compact

When a StateNode has payload.compact set:

{
  "type": "state",
  "payload": {
    "role": "coder",
    "meta": { ... },
    "compact": "<cas_hash_of_summary>",
    "timestamp": 1234
  },
  "refs": [...]
}

This means: "everything before this node has been summarized into the blob at compact". When building LLM context:

1. Walk back from head
2. If a node has compact, stop walking — use the compact summary + all nodes after it
3. If no compact found, use full chain

This enables long-running threads without unbounded context growth.

GC

Simple mark-and-sweep:

1. Roots: all head and start hashes from threads.json + all history/*.jsonl files
2. Mark: from each root, recursively mark all reachable hashes via refs[] (including content Merkle → artifact refs)
3. Sweep: delete unmarked CAS blobs

No per-row format parsing needed. GC only needs to understand refs[].

refs[] Derivation

refs[] is auto-derived from payload at write time via a collectRefs(payload) function that extracts all CAS hash strings from named fields (start, content, ancestors, compact). Application code never reads refs[] — it reads named payload fields. This makes refs[] a pure GC optimization with zero semantic coupling.

Extract Phase

The Extractor is expanded from the current design. Currently it only extracts meta from agent output. In the new design it extracts:

Output	Purpose
meta	Structured metadata (same as before)
contentPayload	The text payload for the content Merkle node
refs[]	CAS hashes of artifacts produced by this role step

The refs[] become the content Merkle node's refs, enabling GC to trace all role-produced artifacts.

What Stays Unchanged

• .info.jsonl — debug logging stays as-is (high-frequency append, not suitable for CAS)
• CAS blob storage format (~/.uncaged/workflow/cas/)
• Bundle registry (workflow.yaml)

Migration

Breaking change. Old .data.jsonl files become incompatible. No backward compat fallback (per project convention).

Changes by Package

Package	Changes
workflow-protocol	Replace StartStep, RoleStep types with StartNode, StateNode. Add ContentMerkleNode type. Expand ExtractResult to include refs[].
workflow-cas	Add findReachableHashes(roots) for GC mark phase
workflow-execute	Rewrite engine to write CAS nodes + update threads.json instead of appending JSONL. Move completed threads to history/. Simplify gc.ts. Simplify fork-thread.ts. Expand extract phase to produce refs.
workflow-runtime	ThreadContext built by walking chain from head. start.prompt resolved from CAS via StartNode.refs[0].
cli-workflow	thread list/show/rm read from threads.json + history/. SSE watches threads.json.
workflow-dashboard	Watch threads.json instead of .data.jsonl
Templates & Agents	Update extract definitions to produce refs[]. Update ctx.start.content → CAS resolved.