Retentioneering Contributing Agent Skill

Contributing to retentioneering-tools

Objective

Convert a user's observation — a bug, a paper cut, a missing capability, a workaround

they keep re-writing — into the smallest upstream change that would have prevented it,

packaged so maintainers can accept it quickly.

Bundled references

| File | Read it when |

|---|---|

| references/repo-conventions.md | before touching code — build/test/docs commands, architecture rules, naming, sync obligations |

| references/proposal-templates.md | when drafting — issue/feature/PR templates with worked examples |

The full route: idea → merged PR

Stage 1 — Capture the observation properly (do this even for "small" ideas)

Record four things while they are fresh:

• Expectation — what the user believed would happen (quote the docstring/docs page

that created the expectation, if any).

• Reality — what actually happened (exact error text or wrong output).
• Cost — time lost, wrong conclusion nearly shipped, workaround written.
• Environment — retentioneering.__version__, Python, OS, install source

(pip wheel vs source checkout).

Field lesson: reports formatted as expectation/reality/cost/repro get acted on;

"X is broken" reports stall.

Stage 2 — Validate against the CURRENT version

Many pain points are already fixed on v5-migration — verify before drafting:

• git log --oneline -30 and CHANGELOG.md — search keywords from the observation.
• Search existing issues/PRs: gh issue list --search "<keywords>", gh pr list ....
• Reproduce on the current checkout (see Stage 3). If it no longer reproduces, the

contribution may become a docs clarification or a regression test instead — both welcome.

Stage 3 — Minimal reproduction (the heart of a bug report)

Build the smallest toy that shows the gap, e.g.:

import pandas as pd
from retentioneering import Eventstream
df = pd.DataFrame({"user_id": ["u1","u1","u2"], "event": ["a","b","a"],
                   "timestamp": pd.date_range("2026-01-01", periods=3, freq="1min")})
# EXPECTED: ...       ACTUAL: ...

Rules: synthetic data only (never the user's real log); deterministic (fixed frames, no

randomness without seed); one behavior per repro; assert the expectation so the repro

doubles as a failing test.

Stage 4 — Choose the contribution shape

| Situation | Shape |

|---|---|

| Clear defect with repro | Issue with repro; PR with fix + regression test if user wants to go further |

| Surprising-but-documented behavior | Docs PR (docstring is the source of truth — site pages regenerate from it) |

| Missing capability | Feature issue: use-case first, proposed signature second, evidence third (see templates) |

| Repeated workaround in user's code | Extract as proposed API: show the workaround, its cost, the proposed call replacing it |

| Wrong-conclusion trap (library was silent) | Frame as "missing signal": what the library knew and did not surface; propose the warning/field |

For API proposals, the accepted framing (from templates): problem → evidence of frequency

→ proposed signature → semantics incl. edge cases → acceptance criteria → migration notes.

Stage 5 — Implement (when making a PR, not just an issue)

Read references/repo-conventions.md first. Non-negotiables:

• Fork or branch from master-tracking v5-migration; one logical change per PR.
• make install-dev — installs deps (uv sync + npm install) and wires the git hook

(a one-time-per-clone step) so commits are auto-checked; skip it and commits bypass the hooks

locally and CI's lint job flags the formatting on your PR. Add make build only when

touching widgets/JS.

• Follow naming conventions (ADR-0008): path_col/event_col/timestamp_col/session_col,

start_event/end_event, verb-first processors, noun widgets, <widget>_data twins.

• All DuckDB execution goes through the unified query engine (L1) — no ad-hoc

duckdb.sql with replacement-scan idioms (superseded ADR-0002).

• Add/extend tests next to the code area (tests/...); a bug fix MUST include the

failing-before test from Stage 3.

• Keep the loud-and-helpful error pattern: errors that list valid values/keys are the

house style; silent degradation (dropped rows, empty results without a signal) is an

auto-reject.

• Sync obligations when renaming/changing API: MCP tool layer mirrors Eventstream

method names; JS metric editor consumes the Python metric schema; docstrings feed the

docs site — update all in the same PR (uv run python docs/scripts/render_pages.py).

Stage 6 — Pre-flight and submit

uv run pre-commit run --all-files      # ruff lint+format, gitleaks, hygiene
uv run pytest tests/ -v                # full suite (CI runs 3.11–3.13)
uv run python docs/scripts/render_pages.py   # if docstrings changed

Commit style: imperative, scoped, explaining WHY when non-obvious (see git log for the

house voice). Update CHANGELOG.md under the unreleased/current section for

user-visible changes.

Submit:

git push -u origin <branch>
gh pr create --title "<imperative summary>" --body-file pr_body.md

PR body (template in references/proposal-templates.md): what & why → linked issue →

repro/before-after → tests added → sync checklist (docs/MCP/JS if applicable) →

breaking-change note. CI must pass: lint + test (3.11/3.12/3.13). master is

PR-only; merging does not release (releases are tag-driven by maintainers).

Stage 7 — Follow through

Respond to review within the PR (avoid force-push after review starts; append commits).

If maintainers ask for direction changes, update the issue first, then the code — the

issue is the contract.

Portfolio mode: many observations at once

When the user accumulated a batch (e.g., a journal of friction from a project):

deduplicate → verify each against current version (Stage 2) → rank by

(frequency × silent-failure risk) → file the top 3–5 as separate issues with repros →

offer one PR for the cheapest verified fix to build credibility, referencing the issues

for the rest. Do not open one mega-issue.