Deleting a Second Platform

How Kepler became a surface, not a fork

The Frame

Kepler shipped fast — by quietly rebuilding the platform inside itself

Kepler is Amplifier's native desktop GUI: a Tauri Rust shell, a React/TypeScript frontend, and a Python sidecar. To move fast, it rebuilt the platform's own machinery internally instead of reusing it.

It worked — but only by becoming a duplicate of the platform. That is the tension this deck resolves.

Provider registry — its own provider_registry.py
Config namespace — its own settings.py under ~/.kepler/
Session pipeline — direct amplifier-foundation calls + a 647-line amplifier_hooks.py

The Proof

The approved design named three duplicated subsystems

This isn't asserted — DISTRO_INTEGRATION_DESIGN.md states it verbatim: Kepler has "its own provider management, its own session creation pipeline, its own configuration namespace (~/.kepler/)."

Three copies of platform machinery, documented before a line was changed.

138

lines in provider_registry.py

222

lines in settings.py

647

lines in amplifier_hooks.py

The Complication

A second platform means every distro improvement is redone — or missed

The design named the cost directly: duplicated infrastructure and divergence risk. Because provider registry, session lifecycle, and config were rebuilt in Kepler instead of reused from distro, upstream work — handoff injection, preflight, streaming — had to be re-implemented or was simply left behind.

The code kept diverging. The fix had to make Kepler stop being its own platform.

The Platform Already There

Kepler carried amplifier-distro as a submodule — but barely used it

amplifier-distro is the platform layer that already powers distro's Slack, Voice, and Web Chat surfaces. Kepler declared it as a git submodule yet delegated almost nothing to it.

The target was in the repo the whole time. The refactor just had to lean on it.

Slack bridge — a distro surface
Voice bridge — a distro surface
Web Chat bridge — a distro surface
Submodule → github.com/ramparte/amplifier-distro.git

The Mechanism

A 5-phase plan rerouted every subsystem onto distro instead of rebuilding it

One idea, applied five times: delegate to distro rather than duplicate. Documented in an approved design and executed phase by phase.

With each subsystem delegated, the old copies had nothing left to do.

1Sessions through distro's Bridge
2Providers to distro's PROVIDERS catalog
3Config migrated to distro.yaml
4Onboarding to distro preflight
5Cleanup — delete the dead machinery

The Turn

Once delegated, the duplicated modules had zero importers left

Delegation turned the old machinery into dead weight. Grepping the sidecar found no remaining provider_registry or settings imports — so the modules could simply be deleted.

Nothing consumed them anymore. That made subtraction the payoff.

138

provider_registry.py deleted

222

settings.py deleted

68

events.py deleted

0

importers remaining in sidecar/

The Payoff

The value was subtraction: 606 dead lines out, 104 tests in

One cleanup commit removed 606 lines of dead code — amplifier_hooks.py shrank from 647 to 41 lines. A follow-up added 104 new tests (211 passing total). Kepler is now a surface, not a fork.

Architecturally equal to distro's Slack, Voice, and Web Chat — it consumes the platform.

606

dead lines removed (647 → 41)

104

new tests added

211

tests passing total

The Takeaway

The best refactor often ships negative lines

Stop rebuilding the platform. Consume it. When a surface delegates instead of duplicating, every upstream improvement arrives for free — and the duplicated machinery becomes deletable dead weight.

23 files changed, net-subtractive where it counted — the win was what got removed.

Sources

Research Methodology

Data as of: repository state through 2026-03-24 (latest commit cb00138).

Feature status: Active / DONE — refactor merged across commits 4eae89d..9288bb9.

Primary source: local checkout of ramparte/amplifier-distro-kepler; amplifier-distro consumed as submodule (github.com/ramparte/amplifier-distro.git).

Commands run:

git show 7543dee --stat ; echo $((647-41)) → 606 lines removed (647 → 41)
git show --stat 9288bb9 → 104 new tests, 211 total (104 new + 107 pre-existing)
for f in ...; do git show 9288bb9:sidecar/tests/$f.py | grep -c 'def test_'; done → 31+33+13+8+19 = 104
git show 9f99f03^:sidecar-shared/{provider_registry,settings,events}.py | wc -l → 138, 222, 68
grep -rn 'provider_registry' sidecar/ ; grep -rn 'from settings import' sidecar/ → 0 matches
sed -n '1,60p' docs/DISTRO_INTEGRATION_DESIGN.md → duplicated-infrastructure problem statement + surface framing
grep -n 'Phase' docs/IMPLEMENTATION_TASKS.md ; git log --oneline 4eae89d..9288bb9 → 5-phase plan
cat .gitmodules → amplifier-distro submodule declaration

Primary contributor: Sam Schillace (author of all 11 refactor-arc commits), co-authored by Amplifier per commit trailers.

Gaps: wc -l of the post-commit amplifier_hooks.py blob reports 38 (a trailing-newline/counting artifact); the headline 606 = author-stated 647 − 41, internally consistent. Design-doc line estimates (129/204/542) differ from the authoritative deletion-commit blob measurements (138/222/647) used here. "211 passing" is per commit body plus an independent def test_ grep count; tests were not re-executed this session.