agentleFS
notactuallytreyanastasionotactuallytreyanastasio/deciduous/CLAUDE.md

deciduous

CLAUDE.md171 starsChanged 23 days agoApache-2.0View on GitHub

Decision graph tooling for AI-assisted development. Track every goal, decision, and outcome. Survive context loss. Query your reasoning. Every system has two stories: Now mode maps current design as decisions. History mode captures evolution and pivots. Take the pulse of a system - what decisions define how it works TODAY. Output: Decision tree of the current model (goal -> options -> decisions). Understand how the system evolved. Narratives are conceptual, not tied to commits. Output: .deciduous/narratives.md with evolution stories. Transform…

  • Reads credentials
  • Installs packages
  • Commits and pushes

The file

Show its rules and text

We split this CLAUDE.md into the rules it gives an agent

19 rules. About 0 are common, 11 are distinctive, and 8 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈11

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 8

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Every system has two stories:

    line 9
  2. Distinctive:

    Look at the current system

    line 44
  3. Distinctive:

    Ask "how did this get this way?"

    line 45
  4. Very specific:

    All slash commands are bootstrapped by deciduous init and updated by deciduous update.

    line 70
  5. Distinctive:

    When a design approach is abandoned and replaced:

    line 87
  6. Distinctive:

    Rule of thumb: If a node describes something the user would put on a project timeline or in a PR description, log it.

    line 245
  7. Very specific:

    Do NOT aggressively prompt for documents: Only suggest when files are directly relevant to a decision node.

    line 279
  8. Distinctive:

    Prompts must be the EXACT user message, not a summary: When a user request triggers new work, capture their full message word-for-word.

    line 283
  9. Very specific:

    When more than one agent works at once, coordinate through deciduous, never through a scratch or markdown file.

    line 432
  10. Very specific:

    Every new session or after context recovery, run /recover or:

    line 454
  11. Very specific:

    When working on this project, identify which domain the work belongs to and use the appropriate subagent context.

    line 500
  12. Distinctive:

    When spawning a Task for exploration or implementation:

    line 514
  13. Very specific:

    After modifying any web/src/ files:

    line 595
  14. Distinctive:

    Ask the user before any destructive operation

    line 709
  15. Distinctive:

    Do NOT commit if tests fail.

    line 740
  16. Distinctive:

    Do NOT commit code that doesn't compile.

    line 746
  17. Distinctive:

    Only commit if ALL pass.

    line 785
  18. Very specific:

    If graphviz is not installed, deciduous dot --png will fail with a helpful error message.

    line 828
  19. Very specific:

    When you run deciduous init, a GitHub workflow is created at .github/workflows/cleanup-decision-graphs.yml.

    line 834

≈ “Common” is an estimate: at least 1 in 5 of the 27,641 CLAUDE.md files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
CLAUDE.md

Deciduous - Decision Graph Tooling

Decision graph tooling for AI-assisted development. Track every goal, decision, and outcome. Survive context loss. Query your reasoning.


The Two Modes

Every system has two stories:

ModeQuestionSkill
Now"How does this work?"/pulse
History"How did we get here?"/narratives → /archaeology

Now mode maps current design as decisions. History mode captures evolution and pivots.


Skills Overview

/pulse - Map Current Design

Take the pulse of a system - what decisions define how it works TODAY.

# Start with a goal
deciduous add goal "API rate limiting behavior" -c 90

# Map options (possible approaches from the goal)
deciduous add option "Identify users by auth token" -c 85
deciduous link 1 2 -r "possible_approach"

deciduous add option "Use IP-based rate limiting" -c 85
deciduous link 1 3 -r "possible_approach"

Output: Decision tree of the current model (goal -> options -> decisions).

/narratives - Understand Evolution

Understand how the system evolved. Narratives are conceptual, not tied to commits.

  1. Look at the current system
  2. Ask "how did this get this way?"
  3. Infer narratives from the design
  4. Find evidence (commits, PRs, docs)
  5. Identify pivots - where the model changed

Output: .deciduous/narratives.md with evolution stories.

/archaeology - Structure for Query

Transform narratives into a queryable graph.

Narrative ElementGraph Node
Titlegoal
Possible approachoption
Choosing an approachdecision
What was learnedobservation
PIVOTrevisit

Output: Connected graph with Now ← revisit ← History.


Slash Commands

All slash commands are bootstrapped by deciduous init and updated by deciduous update.

CommandPurpose
/decisionManage decision graph - add nodes, link edges, sync
/recoverRecover context from decision graph on session start
/workStart a work transaction - creates goal node before implementation
/documentGenerate comprehensive documentation for a file or directory
/build-testBuild the project and run the test suite
/serve-uiStart the decision graph web viewer
/decision-graphBuild a decision graph from commit history (archaeology)
/syncMulti-user sync - pull events, rebuild, push

The Revisit Node

When a design approach is abandoned and replaced:

[Old Decision] ──► [Observation: why it failed] ──► [REVISIT] ──► [New Decision]

The revisit captures:

  • WHAT is being reconsidered
  • WHY (linked observations)
  • Connects old approach to new
deciduous add observation "JWT too large for mobile" -d "Mobile clients hit payload size limits with full JWT claims. Token refresh adds 2-3s on slow connections."
deciduous add revisit "Reconsidering token strategy"
deciduous link <observation> <revisit> -r "forced rethinking"
deciduous status <old_decision> superseded

Node Status

StatusMeaning
activeCurrent truth
supersededReplaced by newer approach
abandonedTried and rejected
# Now mode - only active
deciduous nodes --status active

# History mode - everything
deciduous nodes

# Pivot points
deciduous nodes --type revisit

<!-- deciduous:start -->

Decision Graph Workflow

THIS IS MANDATORY. Log decisions IN REAL-TIME, not retroactively.

Available Slash Commands

CommandPurpose
/decisionManage decision graph - add nodes, link edges, sync
/recoverRecover context from decision graph on session start
/workStart a work transaction - creates goal node before implementation
/documentGenerate comprehensive documentation for a file or directory
/build-testBuild the project and run the test suite
/serve-uiStart the decision graph web viewer
/decision-graphBuild a decision graph from commit history
/syncCheck this copy against the graph server and refresh it

Available Skills

SkillPurpose
/pulseMap current design as decisions (Now mode)
/narrativesUnderstand how the system evolved (History mode)
/archaeologyTransform narratives into queryable graph

The Node Flow Rule - CRITICAL

The canonical flow through the decision graph is:

goal -> options -> decision -> actions -> outcomes
  • Goals lead to options (possible approaches to explore)
  • Options lead to a decision (choosing which option to pursue)
  • Decisions lead to actions (implementing the chosen approach)
  • Actions lead to outcomes (results of the implementation)
  • Observations attach anywhere relevant
  • Goals do NOT lead directly to decisions -- there must be options first
  • Options do NOT come after decisions -- options come BEFORE decisions
  • Decision nodes should only be created when an option is actually chosen, not prematurely

The Core Rule

BEFORE you do something -> Log what you're ABOUT to do
AFTER it succeeds/fails -> Log the outcome
CONNECT immediately -> Link every node to its parent
AUDIT regularly -> Check for missing connections

Through the MCP tools, create and link in one call: add_node with parent_id set to the node it belongs under. Never send add_edge in the same batch as the add_node whose id it needs: the id does not exist yet, and whatever stands in for it (a placeholder, a guess) is refused.

Behavioral Triggers - MUST LOG WHEN:

TriggerLog TypeExample
User asks for a new featuregoal with -p"Add dark mode"
Exploring possible approachesoption"Use Redux for state"
Choosing between approachesdecision"Choose state management"
About to write/edit codeaction"Implementing Redux store"
Something worked or failedoutcome"Redux integration successful"
Notice something interestingobservation"Existing code uses hooks"
Work lands, is dropped, or is replacedstatusdeciduous status 12 completed

Close What You Open

A node that is never closed is listed as open for good. Set its status in the same turn as the work, the way you link a node in the same turn you create it.

StatusMeaning
pendingNot started. A goal, option, decision, action or revisit is born with it
activeIn flight now. Keep it to a handful
completedDone. An outcome or observation is born with it: they record what already happened
rejectedAn option not taken
supersededReplaced by a later node; link that node
abandonedDropped without being done

Open is pending or active. Everything else is closed.

deciduous nodes --open    # what is still open (--status <s>, --json)
deciduous pulse           # open goals, each with its open work
deciduous audit           # what drifted: records left open, committed actions never closed, orphans
deciduous why src/db.rs   # the decisions behind a file (or src/db.rs:120-180): what was chosen, what was rejected

Edge types say what a link means; leads_to is the fallback, not the answer to everything:

  • option -> decision: -t chosen or -t rejected
  • one thing waiting on another: -t blocks; one thing making another possible: -t enables

What NOT to Log - CRITICAL

The decision graph records the USER'S project decisions, not your internal process.

Nodes should capture what the user is building, choosing, and accomplishing. Do NOT create nodes for your own thinking, planning, or tooling steps.

DO NOT create nodes for:

  • Reading/exploring the codebase ("Analyzing project structure", "Reading config files")
  • Your planning process ("Planning implementation approach", "Evaluating options internally")
  • Tool usage ("Running tests to check status", "Checking git log")
  • Context gathering ("Understanding existing auth code", "Reviewing PR comments")
  • Meta-commentary ("Starting work on this task", "Preparing to implement")

DO create nodes for:

  • What the user asked for (goals)
  • Concrete approaches being considered (options)
  • Choices made between approaches (decisions)
  • Code being written or changed (actions)
  • Results of implementation (outcomes)
  • Technical findings that affect decisions (observations)

Rule of thumb: If a node describes something the user would put on a project timeline or in a PR description, log it. If it describes your internal process of reading and thinking, don't.

Document Attachments

Attach files (images, PDFs, diagrams, specs, screenshots) to decision graph nodes for rich context.

# Attach a file to a node
deciduous doc attach <node_id> <file_path>
deciduous doc attach <node_id> <file_path> -d "Architecture diagram"
deciduous doc attach <node_id> <file_path> --ai-describe

# List documents
deciduous doc list              # All documents
deciduous doc list <node_id>    # Documents for a specific node

# Manage documents
deciduous doc show <doc_id>     # Show document details
deciduous doc describe <doc_id> "Updated description"
deciduous doc describe <doc_id> --ai   # AI-generate description
deciduous doc open <doc_id>     # Open in default application
deciduous doc detach <doc_id>   # Soft-delete (recoverable)
deciduous doc gc                # Remove orphaned files from disk

When to suggest document attachment:

SituationAction
User shares an image or screenshotAsk: "Want me to attach this to the current goal/action node?"
User references an external documentAsk: "Should I attach a copy to the decision graph?"
Architecture diagram is discussedSuggest attaching it to the relevant goal node
Files not in the project are dropped inAttach to the most relevant active node

Do NOT aggressively prompt for documents. Only suggest when files are directly relevant to a decision node. Files are stored in .deciduous/documents/ with content-hash naming for deduplication.

CRITICAL: Capture VERBATIM User Prompts

Prompts must be the EXACT user message, not a summary. When a user request triggers new work, capture their full message word-for-word.

BAD - summaries are useless for context recovery:

# DON'T DO THIS - this is a summary, not a prompt
deciduous add goal "Add auth" -p "User asked: add login to the app"

GOOD - verbatim prompts enable full context recovery:

# Use --prompt-stdin for multi-line prompts
deciduous add goal "Add auth" -c 90 --prompt-stdin << 'EOF'
I need to add user authentication to the app. Users should be able to sign up
with email/password, and we need OAuth support for Google and GitHub. The auth
should use JWT tokens with refresh token rotation.
EOF

# Or use the prompt command to update existing nodes
deciduous prompt 42 << 'EOF'
The full verbatim user message goes here...
EOF

When to capture prompts:

  • Root goal nodes: YES - the FULL original request
  • Major direction changes: YES - when user redirects the work
  • Routine downstream nodes: NO - they inherit context via edges

Updating prompts on existing nodes:

deciduous prompt <node_id> "full verbatim prompt here"
cat prompt.txt | deciduous prompt <node_id>  # Multi-line from stdin

Prompts are viewable in the web viewer.

CRITICAL: Maintain Connections

The graph's value is in its CONNECTIONS, not just nodes.

When you create...IMMEDIATELY link to...
outcomeThe action that produced it
actionThe decision that spawned it
decisionThe option(s) it chose between
optionIts parent goal
observationRelated goal/action
revisitThe decision/outcome being reconsidered

Root goal nodes are the ONLY valid orphans.

Quick Commands

deciduous add goal "Title" -c 90 -p "User's original request"
deciduous add action "Title" -c 85
deciduous link FROM TO -r "reason"  # DO THIS IMMEDIATELY!
deciduous serve   # View live (auto-refreshes every 30s)
deciduous remote status   # What this copy and the graph server disagree on

# Metadata flags
# -c, --confidence 0-100   Confidence level
# -p, --prompt "..."       Store the user prompt (use when semantically meaningful)
# -f, --files "a.rs,b.rs"  Associate files
# -b, --branch <name>      Git branch (auto-detected)
# --commit <hash|HEAD>     Link to git commit (use HEAD for current commit)
# --date "YYYY-MM-DD"      Backdate node (for archaeology)

# Branch filtering
deciduous nodes --branch main
deciduous nodes -b feature-auth

CRITICAL: Link Commits to Actions/Outcomes

After every git commit, link it to the decision graph!

git commit -m "feat: add auth"
deciduous add action "Implemented auth" -c 90 --commit HEAD
deciduous link <goal_id> <action_id> -r "Implementation"

The --commit HEAD flag captures the commit hash and links it to the node. The web viewer will show commit messages, authors, and dates.

Branch-Based Grouping

Nodes are auto-tagged with the current git branch. Configure in .deciduous/config.toml:

[branch]
main_branches = ["main", "master"]
auto_detect = true

Audit Checklist (Before Ending a Session)

  1. Does every outcome link back to what caused it?
  2. Does every action link to why you did it?
  3. Any dangling outcomes without parents?

Git Staging Rules - CRITICAL

NEVER use broad git add commands that stage everything:

  • ❌ git add -A - stages ALL changes including untracked files
  • ❌ git add . - stages everything in current directory
  • ❌ git add -a or git commit -am - auto-stages all tracked changes
  • ❌ git add * - glob patterns can catch unintended files

ALWAYS stage files explicitly by name:

  • ✅ git add src/main.rs src/lib.rs
  • ✅ git add Cargo.toml Cargo.lock
  • ✅ git add .claude/commands/decision.md

Why this matters:

  • Prevents accidentally committing sensitive files (.env, credentials)
  • Prevents committing large binaries or build artifacts
  • Forces you to review exactly what you're committing
  • Catches unintended changes before they enter git history

Session Start Checklist

deciduous remote pull     # Refresh the local cache from the graph server
deciduous remote status   # Anything this copy has that the server lacks?
deciduous check-update    # Update needed? Run 'deciduous update' if yes
                          # (auto-checked every 24h if auto-update is on)
deciduous nodes           # What decisions exist?
deciduous edges           # How are they connected? Any gaps?
deciduous doc list        # Any attached documents to review?
deciduous transcript list # Sessions not yet in the graph? import them
deciduous board read --unanswered <label>  # parallel agents: anything waiting for you?
git status                # Current state

The Transcript Is the Log

Logging in real time is the rule, and the transcript is the backstop. Claude Code writes every prompt, edit, command and result of a session to a JSONL file; deciduous transcript import derives from it a goal for each prompt that changed something (the prompt verbatim, already completed: a prompt that led to no edit, commit or test run is not a goal), actions (edit runs, commits) and outcomes (test and build runs with their exit, the closing report), each node carrying metadata.evidence as proven (a tool call and its result) or inferred (the assistant said so). Importing twice adds nothing.

deciduous transcript list                 # this project's sessions: imported or new
deciduous transcript import <session-id>  # or a path to the .jsonl; --dry-run to see counts

When a session ends, a SessionEnd hook in .claude/settings.json runs the import for it ([transcript] auto_import = false in .deciduous/config.toml turns that off). What it derives for a prompt hangs under the goal you logged in that turn (found by the id add goal printed, or by the prompt verbatim: log the goal with -p, and the two meet), and it derives no action where you logged an action in the turn and no outcome where you logged an outcome. A prompt that only says to go on ("ok do it", "sure") is not a goal: its work belongs to the prompt before it.

What the import cannot derive is a decision or an option: whether one approach was chosen over another is a judgement about prose. Log those as you make them. For the ones nobody logged, deciduous transcript decisions <session-id> has a model read the session: each decision it finds is proposed with a quote from the transcript (one whose quote is not there is dropped) and waits in deciduous proposals for accept <n> or reject <n>.

Parallel Agents: The Message Board

When more than one agent works at once, coordinate through deciduous, never through a scratch or markdown file. A file has no ids to answer, no way to ask what is waiting for you, and it is gone with the worktree it was written in.

  • Post interface changes, questions and answers: post_message (MCP) or deciduous board post --as <your-label> -s "subject" -m "body". Address other agents with @label.
  • Read what is waiting for you: read_messages with unanswered_for: "<your-label>", or deciduous board read --unanswered <your-label>. Do it at start, before touching a file another agent may be changing, and before finishing.
  • Reply with reply_to: <id> (--reply-to <id>). A reply is what takes a question off the asker's list; a new post does not.

One board serves every git worktree of the repository, or the server when the project has a [remote]. Messages are not graph nodes: not in graph.json, not exported, not synced. With DECIDUOUS_AGENT_LABEL set in a session's environment, board post uses it for --as, and the board-mentions.sh hook shows that label's unanswered messages at session start and as new ones arrive.

The Graph Server

This project has a [remote] in .deciduous/config.toml: its graph lives on a shared server, and .deciduous/deciduous.db is this machine's cache of it. There is no .deciduous/graph.json to commit or merge, and deciduous sync is not the step here. Every CLI write is queued in .deciduous/remote-log.jsonl and sent before the command exits; MCP tools write to the server directly (pass the workspace named under [remote]).

deciduous remote status   # what differs between this copy and the server (exit 1 if anything)
deciduous remote pull     # send what is waiting, then refresh the local cache from the server
deciduous remote push     # send writes still waiting in the log (the server was down)

Local node ids differ per machine. To point at someone else's node, use its change_id prefix (the CHANGE column in deciduous nodes) or the server id an agent quotes: deciduous link a1b2c3d4 <id> -r "...".

<!-- deciduous:end -->

Session Start Checklist

Every new session or after context recovery, run /recover or:

deciduous check-update    # Update needed? Run 'deciduous update' if yes
                          # (always-on, checks every 24h)
deciduous nodes           # What decisions exist?
deciduous edges           # How are they connected?
deciduous commands        # What happened recently?
git log --oneline -10     # Recent commits
git status                # Current state

Quick Reference

# Build
cargo build --release

# Run tests
cargo test

# Initialize in a new project
deciduous init
# Update Claude integration files to latest version
deciduous update

# Start graph viewer
deciduous serve --port 3000

# Export graph
deciduous sync
deciduous graph > graph.json

# Generate DOT visualization
deciduous dot --png -o docs/decision-graph.dot

# Generate PR writeup
deciduous writeup -t "Feature X" --nodes 1-15 -o PR-WRITEUP.md

Subagents - Domain-Specific Context

Use subagents to scope work to specific parts of the codebase.

When working on this project, identify which domain the work belongs to and use the appropriate subagent context. Subagent definitions are in .claude/agents.toml.

Available Subagents

AgentDomainKey Files
rust-coreCLI, database, export/syncsrc/main.rs, src/db.rs, src/export.rs
webReact/TypeScript viewerweb/src/**/*.{ts,tsx}
toolingClaude Code configuration.claude/, CLAUDE.md, src/init/
docsDocumentation, guidesdocs/, README.md, ROADMAP.md
ciBuild, Actions, releases.github/workflows/, scripts/

How to Use Subagents

When spawning a Task for exploration or implementation:

  1. Identify the domain from the file patterns in .claude/agents.toml
  2. Include the subagent context in your Task prompt
  3. Scope file searches to the relevant patterns

Example: For web viewer work, spawn an Explore agent with:

"Focus on web/src/. This is the web agent domain - React components, D3/Dagre graphs. See .claude/agents.toml for full context."

Why Subagents Matter

  • Reduced context overhead: Focus on relevant files only
  • Domain expertise: Each agent has specialized instructions
  • Parallel work: Multiple agents can work on different domains simultaneously
  • Consistency: Same patterns applied across similar work

Architecture

src/
├── main.rs              # CLI entry, command dispatch
├── lib.rs               # Public API exports
├── db.rs                # SQLite via Diesel: nodes, edges, documents, themes, the command log
├── schema.rs            # Diesel table definitions
├── records.rs           # .deciduous/graph.json: records, reconcile, the merge driver
├── oplog.rs             # .deciduous/remote-log.jsonl: CLI writes queued for the server
├── remote.rs            # deciduous remote: setup, login, status, push, pull, watch
├── server.rs            # Installing and running this machine's server as a user service
├── local_server.rs      # The server's PostgreSQL on localhost:5432
├── transcript.rs        # deciduous transcript: the graph derived from a Claude Code transcript
├── decisions.rs         # transcript decisions / proposals: decisions a model read out of a transcript
├── archaeology.rs       # pivot, timeline, supersede
├── pulse.rs, narratives.rs
├── why.rs               # deciduous why: a file's commits -> their nodes -> the decisions above
├── board.rs             # The message board (deciduous board)
├── mcp/                 # deciduous mcp: the local stdio MCP server (34 tools)
├── api.rs               # deciduous serve --api: the multi-graph JSON daemon
├── serve.rs, viewer.html # deciduous serve: the web viewer, embedded
├── export.rs            # DOT export and PR writeups
├── init/                # deciduous init / update: templates, remote templates, the guard
│   └── known_templates.rs   # @generated: sha256 of every template ever shipped
├── hooks.rs, opencode.rs, config.rs, changelog.rs, github.rs, roadmap.rs
├── events.rs, watch.rs  # Server events over WebSocket
└── demo_swarm/          # deciduous demo-swarm

web/                     # The web viewer: React + TypeScript + D3, one App.tsx
├── src/App.tsx
├── src/types/           # graph.ts, archaeology.ts, qa.ts, generated/ (from the Rust types)
└── dist/                # Built output, a single HTML file

deciduous_mcp/           # The shared graph server: Elixir, PostgreSQL, 21 MCP tools

.deciduous/
├── config.toml          # [remote] names the server and workspace; [branch] controls tagging
├── deciduous.db         # SQLite: with a [remote], this machine's cache; without, the graph
├── remote-log.jsonl     # Writes waiting for the server
├── graph.json           # The git-tracked graph for a project with no server
├── documents/           # Attached files, content-hash named
└── narratives.md        # Narrative tracking

Web Viewer Development

When modifying web viewer code (TypeScript/React), you MUST rebuild and update the embedded HTML.

Key Files

FilePurpose
web/src/App.tsxThe viewer: narrative modes, tree and detail panel, D3 graph, search, themes, the Ask panel
web/src/types/graph.tsTypeScript interfaces for nodes, edges, themes
web/src/types/generated/Types generated from the Rust side; scripts/validate-types.sh checks they match
src/viewer.htmlEmbedded viewer served by deciduous serve
docs/demo/index.htmlStatic demo viewer served on deciduous.dev

Rebuild Process

After modifying any web/src/** files:

# 1. Build the web viewer (outputs singlefile HTML)
cd web && npm run build && cd ..

# 2. Copy to embedded locations (use absolute paths)
cp /path/to/deciduous/web/dist/index.html /path/to/deciduous/src/viewer.html
cp /path/to/deciduous/web/dist/index.html /path/to/deciduous/docs/demo/index.html

# 3. Run Rust tests to ensure nothing broke
cargo test

# 4. Build release binary
cargo build --release

The viewer has no tests of its own (npm test finds no test files): check a change by running deciduous serve and looking.

CLI Commands

CommandDescription
deciduous initInitialize deciduous in current directory
deciduous updateUpdate Claude integration files to latest version
deciduous check-updateCheck if integration files need updating
deciduous auto-update on|offThe once-a-day crates.io version check
deciduous add <type> "title"Add a node (goal/decision/option/action/outcome/observation/revisit)
deciduous link <from> <to>Create edge between nodes
deciduous status <id> <status>Update node status
deciduous nodesList all nodes
deciduous edgesList all edges
deciduous graphOutput full graph as JSON
deciduous commandsRecent CLI writes, with exit status
deciduous transcript listThis project's Claude Code sessions: imported or new
deciduous transcript import <path|session-id>Derive the graph from a transcript, with evidence on every node (--dry-run, --no-hash, --json)
deciduous transcript decisions <path|session-id>A model reads the session and proposes its decisions, each with a quote (--accept, --again, --model)
deciduous proposals [accept|reject <n>]Proposed decisions waiting for review
deciduous backupCreate database backup
deciduous serveStart web viewer
deciduous syncReconcile .deciduous/graph.json with the local DB (projects with no server)
deciduous dotExport graph as DOT format
deciduous writeupGenerate PR writeup markdown
deciduous sync --checkReport pending sync changes, exit 1 if any
deciduous migrateAdd change_id columns for sync
deciduous doc attach <node_id> <file>Attach file to a decision node
deciduous doc list [node_id]List attached documents (all or per-node)
deciduous doc show <id>Show document details
deciduous doc describe <id> [text]Set/update document description (--ai for AI)
deciduous doc open <id>Open document in default application
deciduous doc detach <id>Soft-delete a document attachment
deciduous doc gcGarbage collect orphaned document files
deciduous remote setup|status|push|pull|watchThe shared graph server
deciduous board post|read|showThe message board for parallel agents
deciduous archaeology pivot|supersede|timelineRevisit a decision that turned out wrong
deciduous themes, deciduous tagThemes and tags
deciduous auditAssociate commits with nodes
deciduous why <file>[:lines]The decisions behind a file, through the commits that touched it

DOT Export Options

deciduous dot [OPTIONS]

Options:
  -o, --output <FILE>     Output file (default: stdout)
  -r, --roots <IDS>       Root node IDs for BFS traversal (comma-separated)
  -n, --nodes <SPEC>      Specific node IDs or ranges (e.g., "1-11" or "1,3,5-10")
  -t, --title <TITLE>     Graph title
      --rankdir <DIR>     Graph direction: TB (top-bottom) or LR (left-right)
      --png               Generate PNG file (requires graphviz installed)

Writeup Options

deciduous writeup [OPTIONS]

Options:
  -t, --title <TITLE>     PR title
  -r, --roots <IDS>       Root node IDs (comma-separated, traverses children)
  -n, --nodes <SPEC>      Specific node IDs or ranges
  -o, --output <FILE>     Output file (default: stdout)
      --png <FILENAME>    PNG file to embed (auto-detects GitHub repo/branch for URL)
      --no-dot            Skip DOT graph section
      --no-test-plan      Skip test plan section

Recommended workflow with --auto:

# 1. Generate branch-specific PNG (avoids merge conflicts!)
deciduous dot --auto --nodes 1-11

# 2. Commit and push
git add docs/decision-graph-*.dot docs/decision-graph-*.png
git commit -m "docs: add decision graph"
git push

# 3. Generate writeup with auto PNG detection
deciduous writeup --auto -t "My PR" --nodes 1-11

# 4. Update PR body
gh pr edit N --body "$(deciduous writeup --auto -t 'My PR' --nodes 1-11)"

The --auto flag generates branch-specific filenames (e.g., docs/decision-graph-feature-foo.png) which prevents merge conflicts when multiple PRs each have their own graph.

Database Rules

CRITICAL: NEVER delete the SQLite database (.deciduous/deciduous.db)

The database contains the decision graph. If you need to clear data:

  1. deciduous backup first
  2. Ask the user before any destructive operation

Sharing the graph

This repository has a [remote] in .deciduous/config.toml: the graph lives on the shared server, .deciduous/deciduous.db is this machine's cache of it, and MCP tools write to the server directly. CLI writes are queued in .deciduous/remote-log.jsonl and sent before the command exits.

deciduous remote status   # what this copy and the server disagree on; exit 1 if anything
deciduous remote pull     # send what is waiting, then refresh the cache
deciduous remote push     # resend writes the server did not get (it was down)
deciduous remote watch    # one line per write as it lands, from every agent

A project with no server shares its graph through the git-tracked .deciduous/graph.json instead: deciduous sync reconciles it with the local database both ways (newer updated_at wins, deletes are tombstones), and the deciduous merge driver registered by init/update merges concurrent edits record by record. sync has not exported docs/graph-data.json since 1.0.5; the copy in this repository is a hand-made export from 2026-09-04 that the static demo reads.

Node references anywhere: local id or change_id prefix (deciduous link a1b2c3d4 58). .deciduous/patches/, deciduous diff and .deciduous/sync/ are from 0.16 and 0.17 and do not exist.

Full design: docs/MULTI_USER_SYNC.md and docs/remote-graph.md.


Development Rules

Code Quality - MANDATORY

  1. ALWAYS run tests before committing:
    cargo test
    Do NOT commit if tests fail.
  2. ALWAYS ensure code compiles:
    cargo build --release
    Do NOT commit code that doesn't compile.
  3. Write tests for new functionality:
    • New commands need tests
    • Bug fixes need regression tests
    • Edge cases need coverage
  4. Run clippy for lints:
    cargo clippy

Git Staging Rules - CRITICAL

NEVER use broad git add commands that stage everything:

  • ❌ git add -A - stages ALL changes including untracked files
  • ❌ git add . - stages everything in current directory
  • ❌ git add -a or git commit -am - auto-stages all tracked changes
  • ❌ git add * - glob patterns can catch unintended files

ALWAYS stage files explicitly by name:

  • ✅ git add src/main.rs src/lib.rs
  • ✅ git add Cargo.toml Cargo.lock
  • ✅ git add .claude/commands/decision.md

Why this matters:

  • Prevents accidentally committing sensitive files (.env, credentials)
  • Prevents committing large binaries or build artifacts
  • Forces you to review exactly what you're committing
  • Catches unintended changes before they enter git history

Pre-Commit Checklist

cargo test              # All tests pass?
cargo build --release   # Compiles cleanly?
cargo clippy            # No warnings?

Only commit if ALL pass.


Release Process - MANDATORY

Semver: a new command or a new field in the graph is MINOR; a fix is PATCH. 1.1.0 (2026-10-24) was the last MINOR.

A release is a tag on main that release.yml builds and publishes. What the workflow checks before it publishes anything, and what to do in order:

  1. Bump both package versions. Cargo.toml and deciduous_mcp/mix.exs must carry the tag's version; the workflow refuses a tag whose two versions disagree (Require both package versions to match the publishing tag). cargo build updates Cargo.lock.
  2. Both changelogs. CHANGELOG.md (dated, Added / Changed / Fixed / Not done) and src/changelog.rs (the RELEASES array check-update shows on upgrade, newest first). 0.16.0 through 0.17.1 were once missing from both.
  3. Test and package locally: cargo test, cargo clippy --all-targets, cargo build --release, cargo publish --dry-run. The toolchain is pinned to 1.92.0 in rust-toolchain.toml and the workflows; bump it on purpose, not by letting stable move.
  4. Server deps: cd deciduous_mcp && mix hex.audit. A Hex advisory against any dependency fails the self-contained server build and release acceptance; 1.0.10 was tagged and never published for this, and 1.1.0 needed bandit, plug_crypto and decimal bumped on release day.
  5. PR, merge, then tag by hand from the merge commit:
    git tag -a vX.Y.Z <merge-sha> -m "vX.Y.Z: ..."
    git push origin vX.Y.Z
    Auto Release also fires on the merge and either pushes the tag with GITHUB_TOKEN or fails on the duplicate; a tag pushed by the bot does not trigger release.yml, one pushed by a person does. If the bot got there first, gh workflow run release.yml -f tag=vX.Y.Z -f publish=true.
  6. Watch the run to the end. release.yml builds five CLI binaries and five server executables, runs release acceptance (an isolated PostgreSQL; the readiness wait was fixed in #319 after it raced on 1.1.0's run), then release, publish-crate (crates.io, secret CARGO_REGISTRY_TOKEN) and update-homebrew (the tap repository; the checked-in Formula/deciduous.rb here is a stale copy). A release looks done when nothing was published: confirm gh release view vX.Y.Z, crates.io, and the tap formula's version.
  7. deciduous update in this repository afterwards, and commit what it changed.

Stacked PRs: never gh pr merge --delete-branch on a PR another open PR is based on. GitHub closes the dependent PR instead of retargeting it, and a closed PR whose base is gone cannot be reopened. Merge without deleting, gh pr edit N --base main for the next one, and dispatch type-check.yml on its branch by hand (the CI workflow only runs for PRs that were opened against main). Delete branches at the end.


External Dependencies

Required at Runtime

DependencyRequired ForInstall
NoneCore functionality-

The deciduous binary is self-contained for core features.

Optional Dependencies

DependencyRequired ForInstall
graphviz--png flag (DOT → PNG)brew install graphviz / apt install graphviz

If graphviz is not installed, deciduous dot --png will fail with a helpful error message.


GitHub Action for PNG Cleanup

When you run deciduous init, a GitHub workflow is created at .github/workflows/cleanup-decision-graphs.yml. This workflow:

  1. Triggers after any PR is merged
  2. Finds decision graph PNG/DOT files
  3. Creates a cleanup branch and removes them
  4. Auto-merges the cleanup PR

This keeps your repo clean of accumulated visualization files while still having nice graphs in PRs.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.