Geode Changelog

GEODE Changelog Convention

File Location

/CHANGELOG.md (project root)

Format

Keep a Changelog 1.1.0 + SemVer 2.0.0

## [X.Y.Z] — YYYY-MM-DD

### Added       ← New features
### Changed     ← Changes to existing features (including breaking)
### Fixed       ← Bug fixes
### Removed     ← Removed features
### Architecture ← Structural decisions (GEODE extension)
### Infrastructure ← CI, build, dependencies (GEODE extension)

Versioning Policy

Post-1.0 default: every authorized routine release is a PATCH — features included. Ordinary changes stay under [Unreleased]; a code landing does not itself request a release or version bump. The 0.99.x patch-train (~330 releases where features, fixes, and refactors all bumped PATCH) continues unchanged as 1.0.x. Textbook SemVer's "feature = MINOR" does NOT apply here.

MAJOR.MINOR.PATCH

PATCH: default for every release — features, fixes, refactors, docs+code
MINOR: operator-declared milestone ONLY (never chosen unilaterally)
MAJOR: operator-declared ONLY (breaking pipeline/State schema)

Before choosing ANY non-patch version

1. rg -n "removed in v" core/ evals/ evolve/ — minors may be pledged in advance to deprecation removals (v1.1.0 is pledged to the legacy [self_improving_loop.petri.*] / [self_improving_loop.mutator] removals in evals/config.py). A pledged number is RESERVED; landing unrelated work under it breaks the pledge.
2. Ask the operator. A minor is a product statement, not a diff size.

Incident: the v1.1.0 mis-stamp (2026-07-17, corrected as v1.0.1)

The Slack Socket Mode landing was stamped v1.1.0 by mechanically applying "New feature = MINOR" from this skill's old text + CLAUDE.md §5. Three gaps compounded:

1. Rule/practice divergence — the written rule said feature=MINOR while actual practice was patch-train; the written rule won because it was the only thing in context.
2. Pledge invisibility — v1.1.0's reservation existed only inside runtime deprecation strings, which no release step consults. The grep step above is the guard.
3. Compaction lock-in — once "restamp 1.1.0" entered a session summary, later steps treated the number as settled instead of a decision still needing operator sign-off.

Mis-stamp correction procedure

A wrong version may be RECLAIMED (renamed in place) only while it exists purely as file stamps. Verify against the REMOTE, not a possibly-stale checkout: git ls-remote --tags origin "vX.Y.Z*" (empty), gh release view vX.Y.Z (errors with not-found), and the exact-version PyPI JSON https://pypi.org/pypi/geode-agent/X.Y.Z/json (404). Once any tag/release/artifact exists, the number is burned — correct forward with the next free number instead. To reclaim: rename the CHANGELOG heading in place with a retraction blurb, restamp the 5 locations + site SoT (sync-stats.mjs + check_llms_version.py --fix)

• uv.lock, and land through the normal PR chain.

Scope Rules — What to Record / What Not to Record

What to Record (Feature-level aggregation)

What Not to Record

Writing Procedure

On Code Change Commits (mandatory every time)

Include CHANGELOG entries in the same commit as the code change. Do not defer to a separate commit after the PR.

1. Determine change type → Add 1-line entry to appropriate [Unreleased] category
2. Regenerate architecture inventory when its inputs change; do not hand-count modules/tests
3. Sync README.md metrics + description + Mermaid visualization when changed
4. Bundle code + CHANGELOG + docs in a single commit

Exception: CHANGELOG entry not required for docs/refactor-only changes.

Consistency Verification Checklist (mandatory before PR)

After writing the CHANGELOG, check the following:

□ Do [Unreleased] entries accurately reflect actual code changes?
□ Do Infrastructure section Test/Module counts match measured values?
□ Are all feature/fix items from merged PR bodies included in [Unreleased]?
□ Do README.md Features table and Mermaid diagrams match code behavior?
□ If .claude/mcp_servers.json changed, was the README MCP section also updated?

On Release

Only an authorized release promotes entries and stamps a version. Follow versioning conventions for source stamps, generated mirrors, and immutable artifacts, and GitFlow for branch/CI/integration steps.

1. Promote [Unreleased] to the chosen version and date; leave a fresh empty [Unreleased] heading above it.
2. Regenerate inventory with scripts/architecture_baseline.py --update when its inputs change. File inventory and collected pytest cases are different measurements; neither is a manually maintained CLAUDE.md counter.
3. Regenerate version and public-doc mirrors through their owning commands; verify the diff contains only intended source and generated changes.

Subsection Guide (within Added)

For large releases, organize Added into functional areas:

### Added

#### Core Pipeline
- ...

#### CLI
- ...

#### Memory System
- ...

Choose an affected current package or user-facing capability; examples:

• Runtime and CLI (core)
• Evaluation (evals)
• Experimental Loop (evolve)
• Tools, memory, and provider integrations

Example

## [Unreleased]

### Added
- `AgenticLoop` — while(tool_use) multi-round execution loop
- `BashTool` — HITL shell command execution with 9 blocked patterns

### Fixed
- Scoring confidence edge case with empty analyst array

Git History → Changelog Conversion

When extracting changelog entries from commit logs:

# Extract only feat/fix commits since last release
git log v0.6.0..HEAD --oneline --grep="^feat\|^fix"

1. feat → Added
2. fix → Fixed
3. refactor (with behavior change) → Changed
4. Commits in the same functional area → Aggregate into one
5. Merge/docs/ci/chore → Skip (record ci under Infrastructure when applicable)