decision-graph

Decision Graph Construction

You are building a deciduous decision graph - a DAG that captures the evolution of design decisions in a codebase.

Use the deciduous CLI (at ~/.cargo/bin/deciduous) to build the graph. Run deciduous commands in the current directory (not inside the source repo).

For git commands to explore commit history, use git -C <repo-path> to target the source repo.

CRITICAL: Only use information from the repository itself (commits, code, comments, tests). Do not use your prior knowledge about the project. Everything must be grounded in what you find in the repo.

Commit Exploration

Use a layered strategy to find all relevant commits:

Layer 1: See all commits. Start with the full list when building narratives.

git log --oneline --after="..." --before="..." -- path/

Layer 2: Keyword expansion. Once you have narratives, search for spelling variations and related terms you might have missed (e.g., "cache" → "caching", "cached", "LRU", "invalidate"). For each key identifier in your narratives, trace its full lifecycle:

• Introduction
• Changes and modifications
• Renames
• Deprecation or removal
• Replacement by other mechanisms
• Becoming stable/public API

If there's a feature flag controlling the feature, search for commits mentioning that flag.

Layer 3: Follow authors. If a narrative has a key author, check their commits ±1 month from known commits. They often work on related things.

DO NOT:

• git log ... | head -100 — NO. You will miss commits in the middle.
• git log ... | tail -200 — NO. Same problem.
• Start with keyword filtering — NO. You'll miss things with unexpected names.

DO:

• See all commits first, filter mentally while building narratives
• Include "remove", "delete", "disable", "deprecate" in keyword searches — removals explain transitions
• Check the commit count first (| wc -l), but then see them all
• Read full commit messages for any commit whose title mentions an identifier or concept relevant to your narrative — you need precise understanding of what happened to each one you care about

Finding the Story

Not every commit matters. Look for commits that change the model - how the system conceptualizes the problem:

• Existing tests modified (contract changing, not just bugs fixed)
• Data structures replaced or reworked
• Heuristics changed significantly
• New abstractions introduced
• API behavior shifts

Skip commits that are pure implementation (same model, different code) or routine fixes that just add tests.

Among model-changing commits, find the spine: what question keeps getting re-answered? What approach keeps getting replaced or refined? That's your central thread - build the graph around it.

Narrative Tracking

Don't build the graph as you explore. First, collect commits into narratives.

Maintain narratives.md as you explore:

1. For each significant commit, read narratives.md
2. Ask: "Does this commit evolve an existing narrative?"
3. If yes: append the commit to that narrative's section
4. If no: add a new narrative section

Example narratives.md:

## Cache Strategy

**Arc:** The service initially hit the database on every request, which worked until traffic spiked and the DB became the bottleneck. An in-memory cache solved the latency problem but created a new one: each server instance had its own cache, so users saw inconsistent data depending on which instance handled their request. The team tried cache invalidation broadcasts, but the complexity exploded. Moving to Redis gave a single source of truth at the cost of network latency - but that latency was still 10x better than the DB, and consistency issues disappeared.

- a1b2c3d: Add in-memory cache
- e4f5g6h: Cache invalidation issues
- i7j8k9l: Switch to Redis

## API Rate Limiting

**Arc:** The API originally had no rate limits, which was fine until a misbehaving client brought down the service with a retry loop. The initial fix was simple per-IP throttling, but this broke legitimate use cases like corporate NATs where thousands of users share one IP. The insight was distinguishing between "sustained abuse" and "burst traffic" - a token bucket algorithm lets clients burst up to a limit while still preventing sustained overload. The final refinement added per-endpoint limits after discovering that the /search endpoint was 100x more expensive than others.

- m1n2o3p: Add basic throttling
- ...

The Arc tells the narrative as a story: what problem started it, what was tried, what went wrong, what insight emerged, and where it ended up. If your arc has gaps ("then somehow we ended up with X"), you're missing commits.

Before building the graph, take a critical pass over narratives.md:

• Merge narratives that are essentially the same evolving thing
• Ensure each narrative clearly explains how one independent piece evolved
• Note where narratives branch from or feed into each other

Hardening Phase

After building initial narratives, harden them to ensure nothing is missed.

Assume the commit list is incomplete. Your first pass probably missed things. Before synthesizing the narrative into a graph, actively search for what's missing. If a commit mentions a new term or concept you hadn't searched for, search for that too. Follow the trail until searches stop turning up new relevant commits.

Step 1: Extract concepts per narrative

For each narrative, list the key concepts/APIs/identifiers and their lifecycle stage:

• Introduced: First appearance of the concept
• Changed: Modifications to behavior or implementation
• Renamed/Deprecated/Removed: End of life or replacement
• Marked stable: Became public API or removed "unstable\_" prefix

Example addition to narrative:

## Cache Strategy
Concepts: cache, LRUCache, cacheTimeout, invalidate

Lifecycle:
- cache: introduced (a1b2c3d), changed (e4f5g6h), renamed to LRUCache (x1y2z3)
- cacheTimeout: introduced (e4f5g6h), removed (i7j8k9l)
- LRUCache: introduced via rename (x1y2z3), marked stable (p1q2r3)

Commits:
- a1b2c3d: Add in-memory cache
- ...

Step 2: Exhaustive search per concept

For each concept, search full commit messages (not just subject lines):

git log --all --after="..." --before="..." --grep="<concept>" --format="%H %s" -- path/

For each match, read the full commit message:

git show <sha> --format="%B" --no-patch

Step 3: Follow replacements

When something is removed or deprecated, trace what replaced it. The replacement mechanism is its own concept that needs lifecycle tracking.

Example: If you find "Remove ManualCache in favor of AutoCache", then:

1. Add AutoCache to your concepts list
2. Search for all commits mentioning AutoCache
3. Track its full lifecycle (introduced, changed, etc.)

Replacements often represent a distinct design phase. Don't collapse "X removed, Y introduced" into a single node - model Y's introduction and evolution as its own chain.

Step 4: Rewrite narratives with gaps filled

Rewrite narratives.md integrating any newly discovered commits. The rewritten version should:

• Include ALL commits found for each concept
• Include replacement mechanisms as first-class concepts
• Update the lifecycle tracking for each concept
• Ensure the arc is complete (if something was introduced, when was it changed/removed?)

If a concept has an incomplete arc (e.g., introduced but never removed, yet it's not in current code), investigate further.

Cross-Narrative Connections

When building the graph, capture how narratives relate - but only connect things that are causally related.

The Causality Test

Before connecting A → B, ask: Would B have happened anyway if A hadn't existed?

• Yes, B would happen anyway → They're parallel concerns. Both branch from their common parent (often the goal).
• No, B only exists because of A → Sequential. A leads_to