---

role: file-system-state-management summary: | File-system state management for OpenProse. The normative reference for the canonical world-model artifact layout, deterministic serialization, the append-only receipt ledger, and the workspace (private scratch) vs published (canonical truth) discipline. see-also:

• ../prose.md: VM execution semantics + the render harness seam
• ../forme.md: Wiring semantics (produces the topology world-model)
• ../primitives/session.md: Session context and compaction guidelines

---

File-System State Management

This document describes how the OpenProse VM tracks execution state using files under <openprose-root>. Native repositories use the repository root as <openprose-root>. Attached repositories use repo/.agents/prose. User-global work uses ~/.agents/prose.

This file is the normative reference for filesystem artifact layout and file formats. prose.md summarizes the same model from the execution algorithm's point of view; when details differ, prefer this file for paths, ownership, and serialization formats.

Overview

File-based state persists all execution artifacts to disk. This enables:

• Inspection: See exactly what happened at each step, including intermediate work
• Resumption: Pick up interrupted runs from the last completed receipt
• Debugging: Trace through the compiled intent, workspace scratch, and the published world-model
• Auditability: The append-only receipt ledger chains every commit; the world-model is content-addressed and diffable

Key principle: Files are inspectable artifacts. The directory structure IS the execution state.

The load-bearing distinction (the invariant state/README.md states for every backend): a node's published world-model is the canonical, deterministically-serialized, fingerprinted artifact — the truth downstreams subscribe to. The render's workspace is private scratch — intermediate reasoning, working notes — and is never fingerprinted and never subscribed to. This is not mere output-visibility (the earlier workspace/→bindings/ distinction): it is fingerprint-materiality. A node updates its published world-model only when something semantically material actually changed; immaterial churn stays in the workspace. SQL/vector indices are derived projections of the canonical artifact, never the truth.

---

Directory Structure

# <openprose-root>
├── src/                                    # Authored OpenProse source
│   ├── research-system/
│   │   ├── index.prose.md                  # Conventional multi-file program root
│   │   ├── researcher.prose.md
│   │   └── synthesizer.prose.md
│   ├── patterns/
│   │   └── worker-critic.prose.md
│   └── tests/
│       └── research-system.test.prose.md
├── dist/                                   # Compiled intent (topology WM + canonicalizers + validators)
│   ├── manifest.next.json                  # Newly compiled, pending activation
│   └── manifest.active.json                # Currently active compiled intent
├── runs/
│   └── {YYYYMMDD}-{HHMMSS}-{random}/
│       ├── compiled-intent.json            # Topology world-model + canonicalizers + validators (snapshot)
│       ├── root.prose.md                   # Copy of the invoked source file
│       ├── sources/                        # Source files copied at compile time
│       │   ├── researcher.prose.md
│       │   ├── critic.prose.md
│       │   └── synthesizer.prose.md
│       ├── workspace/                      # Private render scratch — NEVER fingerprinted, NEVER subscribed to
│       │   ├── researcher/
│       │   │   ├── notes.md                # Intermediate scratch work
│       │   │   ├── raw-results.md          # Intermediate data (e.g. raw poll w/ fetched_at)
│       │   │   └── __delegate/             # Runtime delegation state (if any)
│       │   │       └── {delegate}/
│       │   │           ├── {id}.md          # Request payload
│       │   │           └── {id}-response.md # Response payload
│       │   ├── critic/
│       │   │   └── evaluation.md
│       │   └── synthesizer/
│       │       └── report.md
│       ├── world-model/                    # Published canonical truth (one per node) — deterministically serialized + fingerprinted
│       │   ├── caller/                     # Caller-provided inputs (gateway-style)
│       │   │   └── question.md
│       │   ├── researcher/                 # Researcher's published world-model artifact
│       │   │   ├── findings.md
│       │   │   ├── sources.md
│       │   │   └── .version                # ContentAddress of this committed version
│       │   ├── critic/
│       │   │   ├── evaluation.md
│       │   │   └── .version
│       │   └── synthesizer/
│       │       ├── report.md
│       │       └── .version
│       ├── receipts/                       # Append-only receipt ledger (one chain per node)
│       │   ├── researcher.jsonl            # Receipt chain (prev pointers); status/fingerprints/wake/cost
│       │   ├── critic.jsonl
│       │   └── synthesizer.jsonl
│       ├── vm.log.md                        # Append-only human-readable execution log
│       └── agents/                         # Run-scoped agent memory
│           └── {name}/
│               ├── memory.md
│               ├── {name}-001.md
│               └── ...
├── state/                                  # Durable cross-run state
│   ├── agents/
│   │   └── {name}/
│   │       ├── memory.md
│   │       └── ...
│   └── world-model/                        # Durable per-node canonical truth (survives runs)
│       └── {node}/
│           ├── published/                  # Current committed canonical artifact
│           ├── versions/                   # Content-addressed prior versions ({ContentAddress})
│           └── receipts.jsonl              # The node's durable receipt ledger
├── deps/                                  # Cloned dependency repos (gitignored)
│   ├── github.com/
│   │   ├── openprose/
│   │   │   └── prose/                      # Full clone of github.com/openprose/prose
│   │   │       └── packages/
│   │   │           ├── std/
│   │   │           │   └── evals/
│   │   │           │       └── inspector.prose.md
│   │   │           └── co/
│   │   │               └── systems/
│   │   │                   └── company-repo-checker/
│   │   │                       └── index.prose.md
│   │   └── alice/
│   │       └── research/
│   │           └── ...
│   └── gitlab.com/
│       └── ...
├── prose.lock                             # Pinned dependency SHAs (committed to git)
└── .env                                   # Config (simple key=value format)

# User-global OpenProse root
~/.agents/prose/
├── src/                                    # User/global scoped source
├── runs/                                  # User/global scoped run state
├── state/                                 # User-global durable cross-run state
│   ├── agents/
│   └── world-model/
├── deps/                                  # User/global scoped dependency cache
├── prose.lock
└── .env

Run ID Format

Format: {YYYYMMDD}-{HHMMSS}-{random6}

Example: 20260317-143052-a7b3c9

Segment Numbering

Agent segments use 3-digit zero-padded numbers: captain-001.md, captain-002.md, etc.

---

The Three Directories

The core of Prose state management is the separation of three directories: sources/ (immutable compiled snapshots), workspace/ (private, never-fingerprinted render scratch), and world-model/ (the published, deterministically-serialized, fingerprinted canonical truth).

sources/ — Source Snapshots

*.prose.md files copied at compile time. These are the definitions as they were at compile time — stable snapshots even if source files change during execution.

Written by: the compile phase Read by: The VM when constructing render prompts Immutable during execution.

workspace/ — Private Render Scratch

One subdirectory per node. The render writes all its working state here — intermediate notes, drafts, raw poll data, scratch reasoning.

Written by: the render (each node writes to its own subdirectory) Read by: only the render that owns it, and the VM for __error.md. Never fingerprinted, never subscribed to. Everything is preserved for post-run debugging.

The workspace is the render's private sandbox. Critically, immaterial churn lives and dies here: a re-poll that only bumps a fetched_at timestamp writes to workspace, the canonicalizer drops the immaterial field, the fingerprint does not move, and nothing reaches the published truth. It can contain anything:

workspace/researcher/
├── search-log.md           # What searches were attempted
├── raw-results.md          # Unfiltered search results (may carry fetched_at, request ids)
├── filtered-results.md     # After relevance filtering
└── notes.md                # Scratch thinking

world-model/ — Published Canonical Truth

One subdirectory per node (plus caller/ for external/caller inputs). The node's maintained truth — a content-addressable artifact (a directory by default), the public interface downstream nodes subscribe to.

Written by: the VM via commit_world_model, on a rendered receipt with a moved fingerprint (the canonical-serialization-before-fingerprint pass, below). Read by: downstream renders — by reference, at a pinned content-addressed version (never inlined into context).

world-model/
├── caller/
│   └── question.md         # External/caller input (gateway-style node)
├── researcher/
│   ├── findings.md         # Published, structured, canonicalizable truth
│   ├── sources.md
│   └── .version            # ContentAddress of this committed version
└── synthesizer/
    ├── report.md
    └── .version

Note: report.md here is the structured truth that backs the fingerprint. Free-form rendered prose is a derived projection excluded from the fingerprint (the structured-backing rule in contract-markdown.md) — otherwise an LLM re-rendering the same paragraph hashes differently every time and falsely re-triggers downstreams. Fingerprint the structured truth; render prose from it.

Faceted layout — one subtree per facet

When a node's ### Maintains declares facets by naming the parts (a #### sub-heading inside ### Maintains is a facet — the named-parts rule in contract-markdown.md), the published artifact lays each facet out as its own subtree under the node directory: published/<facet>/…. The directory structure is the subscription surface — the same name is the facet's fingerprint unit, its Requires.<facet> ↔ Maintains.<facet> subscription symbol, and its on-disk region. The canonical competitor-activity-monitor example (examples/competitor-activity/) declares #### funding, #### hiring, and #### product-launches, so its published artifact is:

world-model/
└── competitor-activity-monitor/
    ├── competitors.md          # shared, un-facetted fields (name, last_corroborated)
    ├── published/
    │   ├── funding/            # #### funding facet subtree
    │   │   └── events.md       # structured funding events (round, amount, date)
    │   ├── hiring/             # #### hiring facet subtree
    │   │   └── roles.md        # department set + open-role count
    │   └── product-launches/   # #### product-launches facet subtree
    │       └── launches.md     # launch set + shipped status
    └── .version                # ContentAddress of this committed version

The fingerprinting rule with facets (see the canonical-serialization pass below). The single authority for facet tokens is the compiled canonicalizer that travels with the contract: it reduces the node's structured material truth (the ### Maintains canonicalization spec, frozen at compile, addressing material fields by their dotted structured paths) to the { facet → token } map. The published/<facet>/… directory layout above is a legibility convention for the on-disk artifact — it mirrors the facets so "the directory is the state" reads literally — but the per-facet token is not a digest of that on-disk subtree's bytes; it is what the compiled canonicalizer computes over the facet's material structured content.

• The atomic @atomic token is computed over the whole published/ tree — every facet's material content plus the shared un-facetted fields — and is always emitted. It is the diamond-reconvergence primitive and the free default.
• Each declared facet additionally emits one token computed over only that facet's material content (its #### funding material fields). A downstream that ### Requires funding and resolves to #### funding subscribes to that facet's token: a move in the hiring facet advances the hiring token and the @atomic token but not the funding token, so the funding-only subscriber does not wake (the selector boundary).
• Shared un-facetted fields (here competitors.md's name / last_corroborated) belong to no facet's material content, so they move only the @atomic token.

Atomic-only nodes (no #### parts) keep the flat layout above — facets are purely additive, and the @atomic token over the whole artifact is unchanged. Correctness holds either way: the compiled canonicalizer fingerprints the structured material truth regardless of the on-disk directory shape; the subtree layout is the legibility surface, not a second fingerprinting convention.

---

The Canonical-Serialization-Before-Fingerprint Pass

This is the pass the old workspace/→bindings/ copy lacked. Publishing a node's world-model is not a dumb file copy; it is serialize canonically → canonicalize (drop immaterial) → fingerprint → sign receipt:

1. Collect the node's published artifact files from its commit set.
2. Canonically serialize the artifact deterministically: stable file ordering (sorted by normalized path), path/encoding normalization, and sorted-key JSON for any structured records (reuses the receipt's canonical-JSON machinery). The same byte sequence must result regardless of write order or filesystem enumeration order.
3. Apply the compiled canonicalizer for this node (### Maintains canonicalization spec, frozen at compile time): drop immaterial fields (fetched_at, request ids, cosmetic ordering), normalize sets/numbers/text to declared tolerances. This yields the canonical (material) form.
4. Fingerprint = the compiled canonicalizer's digest over the canonical material form, producing the FingerprintMap — the atomic @atomic token (over the whole material truth) always, plus one token per declared facet (over only that facet's material content; see Faceted layout above). The token authority is the compiled canonicalizer over the structured truth, not a digest of the on-disk published/<facet>/ subtree bytes.
5. Compare against the node's prior receipt fingerprints. If nothing moved, write a skipped receipt (unchanged fingerprints copied forward, empty semantic_diff, zero cost) and publish nothing new. If a fingerprint moved, write the canonical artifact to world-model/{node}/, stamp its .version with the content address, and emit a rendered receipt carrying the new fingerprints.

The reconciler's wake decision is this fingerprint comparison — deterministic, total, no LLM. Only a rendered receipt with a moved fingerprint propagates to downstreams (the reconcile loop in concepts/reconciler.md).

---

File Formats

Compiled Intent

The compile-phase output. For filesystem runs it may be snapshotted as compiled-intent.json. See forme.md and the compiler IR for the full format. Contains:

• the topology world-model (Forme's output): nodes (declared contracts), edges (resolved subscriptions: subscriber.Requires.<facet> → producer.Maintains.<facet>), entry_points, and acyclic: true
• per-node canonicalizers (the frozen ### Maintains canonicalization spec)
• per-node postcondition validators (deterministic | render-attested)
• contract_fingerprints frozen at compile

Written by: the compile phase (Forme + canonicalizer-compile + postcondition-compile) Read by: the run phase (the reconciler resolves propagation targets from edges)

Caller / Gateway Input Files

Path: world-model/caller/{name}.md

# question

binding: input
source: caller

---

What are the latest developments in quantum computing?

Written by: The VM at run start (from CLI args, config, or user prompt)

Run-Typed Inputs

When a requires entry has type run or run[], the VM writes a structured binding with metadata instead of a plain value.

For a single run:

# subject

binding: input
source: caller
type: run

---

run: 20260406-201439-1a3369
path: <openprose-root>/runs/20260406-201439-1a3369
root: customer-discovery
status: complete

For run[]:

# runs

binding: input
source: caller
type: run[]

---

- run: 20260406-201439-1a3369
  path: <openprose-root>/runs/20260406-201439-1a3369
  root: customer-discovery
  status: complete

- run: 20260407-031438-bf26a3
  path: <openprose-root>/runs/20260407-031438-bf26a3
  root: competitive-landscape
  status: complete

The downstream render receives the path and can read the run's published world-model/, receipts/, vm.log.md, and compiled intent directly. The structured header gives the render immediate access to key metadata without traversing the filesystem.

Resolution order for run references:

• Bare ID (e.g., 20260406-201439-1a3369): resolves to <openprose-root>/runs/{id}
• ~/{id}: resolves to ~/.agents/prose/runs/{id} (user scope)
• Absolute path: used as-is

Written by: The VM at binding time (before any render begins)

World-Model Artifact Files

Path: workspace/{node}/{name}.md (private render scratch — never fingerprinted) Path: world-model/{node}/{name}.md (published canonical truth — fingerprinted)

Artifact files are simple Markdown — just the structured content. No special frontmatter required:

# Findings

## Claim 1: Transformer architectures dominate NLP benchmarks
- Source: arxiv.org/abs/1706.03762
- Confidence: 0.95

## Claim 2: Scaling laws predict performance from compute
- Source: arxiv.org/abs/2001.08361
- Confidence: 0.88

Written by: the render (to workspace). The VM commits the published artifact via commit_world_model — serialize canonically → canonicalize → fingerprint → sign — and publishes only when a fingerprint moved.

Error Files

Path: workspace/{node}/__error.md

# Error: no-results

No relevant sources found for the topic.

Searched:
- Google Scholar: 0 relevant results
- arXiv: 2 results, both tangential

Partial data: None available.

The __ prefix signals to the VM that this is an error, not a committed artifact.

Written by: the render (when it cannot satisfy its ### Maintains). Triggers a failed receipt; nothing commits, no fingerprint moves.

---

vm.log.md — Append-Only Execution Log

vm.log.md is an append-only log of execution events. The VM appends entries as execution progresses.

Only the VM writes this file. Subagents never modify vm.log.md.

Format

# run:20260317-143052-a7b3c9 deep-research
upstream: [20260306-112233-f4a5b6]     # optional — present when run has run-typed inputs
root: research/deep-research          # always present — the invoked contract file

1→ [input] question ✓
2→ researcher ✓ rendered
3→ ∥start critic,fact-checker
3a→ critic ✓
3b→ fact-checker ✓
3→ ∥done
4→ synthesizer ✓ rendered
---end 2026-03-17T14:35:22Z

The header is the block between the # heading and the first event marker:

• upstream: is written once at binding time, before any render begins. Omitted when the run has no run-typed inputs.
• root: is always present — the invoked contract file.
• On resumption, the VM reads these as context but does not re-process them.

Event Markers

Marker	Meaning	Example
N→ [input] name ✓	Caller input bound	1→ [input] question ✓
N→ node ✓ rendered	Render committed a moved world-model	2→ researcher ✓ rendered
N→ node ∅ skipped	Reconciler skipped: no fingerprint moved	2→ researcher ∅ skipped
N→ ∥start a,b	Parallel renders started	3→ ∥start critic,fact-checker
Na→ a ✓	Parallel render completed	3a→ critic ✓
N→ ∥done	All parallel renders complete	3→ ∥done
N→ node ✗ error-name	Render signaled an error	3→ researcher ✗ no-results
N→ node ⇒ delegate (delegate: {id})	Render yielded to a runtime delegate	4→ server ⇒ synthesizer (delegate: req-001)
N→ delegate ✓	Runtime delegate completed	4→ synthesizer ✓
N→ node ⟳ (resumed)	Render resumed after delegation	4→ server ⟳ (resumed)
N→ [eval] assertion ✓	Test assertion passed	5→ [eval] assertion ✓
N→ [eval] assertion ✗	Test assertion failed	5→ [eval] assertion ✗
---test PASS	Test passed (all assertions satisfied)	---test PASS
---test FAIL (N/M assertions)	Test failed	---test FAIL (2/3 assertions)
---end TIMESTAMP	Run completed	---end 2026-03-17T14:35:22Z
---error TIMESTAMP msg	Run failed	---error 2026-03-17T... no-results

When the VM Writes

Event	Action
Caller input bound	Append input marker
Render completes	Append completion marker
Parallel starts/joins	Append parallel markers
Error occurs	Append error marker
Delegation spawned	Append ⇒ marker
Delegate completes	Append delegate ✓ marker
Render resumed	Append ⟳ marker
Test assertion evaluated	Append [eval] marker
Test suite result	Append ---test marker
Run ends	Append end marker

The VM does NOT rewrite the entire file. Each write is a single line append.

Resumption

To resume an interrupted run:

1. Read the receipts/ ledger — the append-only chain is the source of truth; find the last committed receipt per node
2. Read the compiled intent — get the topology and propagation edges
3. Scan world-model/ — confirm published canonical artifacts (and their .version)
4. Re-derive reconciler dirty/coalesce state from unconsumed upstream receipts and continue

---

Who Writes What

Artifact	Written By	When
compiled intent	the compile phase	Before execution
root.prose.md	the compile phase	Before execution
sources/*.prose.md	the compile phase	Before execution
world-model/caller/*.md	VM	At entry / gateway boot
workspace/{node}/*	the render	During the render
workspace/{node}/__delegate/{delegate}/{id}.md	the render	Before delegation yield
workspace/{node}/__delegate/{delegate}/{id}-response.md	VM	After delegate completes
world-model/{node}/* + .version	VM (commit_world_model)	On a rendered receipt with a moved fingerprint
receipts/{node}.jsonl	VM	After each render or skip
vm.log.md	VM	After each event
runs/{id}/agents/{name}/memory.md	Execution-scoped agent	During the render
runs/{id}/agents/{name}/{name}-NNN.md	Execution-scoped agent	During the render
state/agents/{name}/memory.md	Durable cross-run agent	During the render
state/agents/{name}/{name}-NNN.md	Durable cross-run agent	During the render
state/world-model/{node}/published/ + versions/ + receipts.jsonl	VM	On durable commit

Key principle: The VM orchestrates. The render writes its working state to its private workspace. The VM publishes the canonical world-model via the canonical-serialization-before-fingerprint pass and signs a receipt — never a dumb copy.

---

The Commit-and-Sign Protocol

This replaces the old copy-on-return for nodes. When a render completes:

1. The render writes all its working state to workspace/{node}/.
2. The render returns a confirmation listing its committed artifact files and attesting its ### Maintains postconditions (no separate judge beat).
3. The VM runs the canonical-serialization-before-fingerprint pass (serialize canonically → apply the compiled canonicalizer → fingerprint).
4. The VM compares the new FingerprintMap against the node's prior receipt.
   • No fingerprint moved → write a skipped receipt (unchanged fingerprints copied forward, empty semantic_diff, zero cost); publish nothing.
   • A fingerprint moved → write the canonical artifact to world-model/{node}/, stamp .version, and emit a rendered receipt with the new fingerprints.
5. The VM appends the receipt to receipts/{node}.jsonl (with its prev pointer) and a marker to vm.log.md.

Publishing is write world-model + sign receipt. Only a rendered receipt with a moved fingerprint propagates to downstreams.

If the render wrote __error.md instead:

1. VM reads workspace/{node}/__error.md
2. VM emits a failed receipt — a failed render commits nothing; the prior world-model stands and no fingerprint moves (failure = no-commit; see concepts/reconciler.md)
3. VM appends error marker to vm.log.md

---

Agent Memory Files

Agent memory lives under runs/{id}/agents/{name}/ for execution-scoped sessions and under state/agents/{name}/ for durable cross-run sessions.

{agent-scope}/{name}/memory.md

The agent's current accumulated state:

# Agent Memory: captain

## Current Understanding

The project is implementing a research pipeline for quantum computing.
Researcher produces good breadth but sometimes lacks depth on subtopics.

## Decisions Made

- 2026-03-17: Approved initial research scope, flagged need for deeper source verification
- 2026-03-17: Set confidence threshold at 0.7 for claim inclusion

## Open Concerns

- Source diversity is low — too many arXiv papers, not enough industry reports

{agent-scope}/{name}/{name}-NNN.md

Prior segment records:

# Segment 001

timestamp: 2026-03-17T14:32:15Z

## Summary

- Reviewed: researcher output (findings.md, sources.md)
- Found: 12 claims extracted, 3 below confidence threshold
- Decided: Accept 9 claims, request broader source search for rejected 3
- Next: Review critic evaluation, verify source diversity improved

Memory Scoping

Scope	Declaration	Path	Lifetime
Execution (default)	### Runtime with persist: true	<openprose-root>/runs/{id}/agents/{name}/	Dies with run
Project	### Runtime with persist: project	<openprose-root>/state/agents/{name}/	Survives runs
User	### Runtime with persist: user	~/.agents/prose/state/agents/{name}/	Survives projects

---

<openprose-root>/.env

Simple key=value configuration:

OPENPROSE_DEFAULT_MODEL=opus
OPENPROSE_MAX_PARALLEL=5

---

Nested Imports

When a run imports and invokes another OpenProse program (via installed dependency or local file), the imported program runs in its own subdirectory:

<openprose-root>/runs/{id}/imports/{handle}--{slug}/
├── compiled-intent.json
├── root.prose.md
├── sources/
├── workspace/
├── world-model/
├── receipts/
├── vm.log.md
├── imports/                    # Further nesting
│   └── ...
└── agents/

Same structure recursively, enabling unlimited nesting depth.

---

Summary

Prose file-system state management separates private scratch from published, fingerprinted truth:

1. sources/ — immutable source snapshots (what was compiled)
2. workspace/ — private render scratch (how a node did its work) — never fingerprinted, never subscribed to
3. world-model/ — the published canonical truth per node (deterministically serialized + fingerprinted) — what downstreams subscribe to
4. receipts/ — the append-only receipt ledger (the source of truth for reconciler state)

The compiled intent defines the topology. The reconciler walks it by comparing fingerprints — deterministic, total, no LLM. The render writes to workspace. The VM commits the canonical world-model through the canonical-serialization-before-fingerprint pass and signs a receipt — never a dumb copy. SQL/vector indices, when present, are derived projections of the canonical truth. Everything is on disk, everything is inspectable, everything is auditable through the signed receipt chain.