The Marathon Session

How one agent shipped a terminal UI

The Setup

Point an agent at an empty repo for days and you expect it to lose the thread

This autonomous Amplifier run did the opposite: it built a full Textual-based terminal UI — the repo ramparte/amplifier-tui — from an empty directory.

The obvious question: how far did it actually get before drifting?

# the repo the marathon session built
$ git remote -v
origin  git@github.com:ramparte/amplifier-tui.git

# first commit — start of the build
bf6d4bc  2026-02-05 20:24  feat: initial
         amplifier-tui project

The Evidence

In one dense ~24-hour burst, it shipped 138 feature commits

Git records 142 commits — 138 of them feat: — between Feb 7 11:42 and Feb 8 12:05. Across the whole repo, 296 commits carry the Co-Authored-By: Amplifier trailer.

So it shipped at scale. But what exactly did all that output become?

142

commits in the ~24h burst

138

feat: commits in that burst

296

Co-Authored-By: Amplifier commits, repo-wide

~24h

continuous burst, after a 38.3h pause

What Shipped

The output is real software you can run today

amplifier-tui is a Textual terminal UI: 55,134 lines of Python, 159 _cmd_ command handlers, and 11 built-in color themes (dark, nord, dracula, and more).

A run that produces this much should not have survived — here's why.

55,134

lines of Python (excl. .venv/.git)

159

_cmd_ command handlers

11

built-in color themes

284

total commits, Feb–Apr 2026

The Problem Everyone Predicts

A run this long should have collapsed

Every context window fills, then gets wiped by compaction. The expected ending of a marathon autonomous run is drift: the agent forgets what it was doing and loses the thread.

So how did this one keep shipping across all those memory resets?

Mechanism · Part 1

The orchestrator never held the whole job

build-session.md documents an "Autonomous Orchestrator" that delegates exploration, design, build, and bug-hunting to disposable foundation:* sub-agents — so no single context has to remember everything.

That handles the work. But where did the orchestrator keep its own memory?

foundation:explorer — investigates the codebase
foundation:zen-architect — designs the approach
foundation:modular-builder — writes the feature
foundation:bug-hunter — finds and fixes defects

Mechanism · Part 2

Its real memory lived on disk, not in context

SESSION-HANDOFF.md and .amplifier/context/build-session.md are git-tracked files. When context is wiped, a resumed session reloads them and picks up the thread — memory that outlives every compaction.

Together these two moves are the whole survival mechanism.

SESSION-HANDOFF.md — 6,952 bytes, added by commit 4a89aaa; project state & architecture for a resuming session
build-session.md — 15,537 bytes, git-tracked; the orchestrator's operating instructions

Why It Held Together

Small orchestrator + disposable workers + memory on disk is why it kept shipping

And it finished clean: a disciplined 6-phase refactor (Phase 0–5, then a Phase 6 pytest suite), with a "Build Verification First" milestone that committed testing infrastructure on its own.

1,454 test functions across 40 files now stand behind the result.

Phase 0–5 — type safety → widgets → commands → persistence → features → error handling
Phase 6 — comprehensive pytest suite (1,454 test functions, 40 files)
Build Verification First — test infra committed as its own milestone (4a89aaa, 8b988ba)

The Pattern To Keep

Long-running agents don't need a bigger memory

They need to delegate to disposable workers and keep their real state in files. A tiny orchestrator plus memory on disk beats trying to hold everything in one ever-growing context.

Small orchestrator. Disposable sub-agents. Memory on disk. That's how a marathon run finishes instead of drifting.

Sources

Research Methodology

Repo state as of: 2026-04-23 (last commit c967de0). Feature status: Shipped.

Primary sources: the ramparte/amplifier-tui git repo, .amplifier/context/build-session.md, SESSION-HANDOFF.md, and README.md.

Commands run:

Repo identity: git remote -v → git@github.com:ramparte/amplifier-tui.git
Burst size: git log --since='2026-02-07 11:42' --until='2026-02-08 12:06' --oneline | wc -l → 142 commits, 138 feat:
Gap analysis: git log --reverse --format='%ct|%ci|%s' | awk → 38.3h gap; ~24h continuous burst
Trailer count: git log --format='%H %s%n%b' | grep -c 'Co-Authored-By: Amplifier' → 296 (repo-wide)
Totals: git log --oneline | wc -l → 284 commits; grep -c 'feat:' → 174 feat:
Line count: find . -name '*.py' -not -path './.venv/*' -not -path './.git/*' | xargs wc -l → 55,134
Commands: grep -rh 'def _cmd_' --include='*.py' . | wc -l → 159
Tests: grep -rh 'def test_' tests/*.py | wc -l → 1,454 across 40 files
Themes: grep -i theme README.md → 11 built-in themes
Architecture: grep -niE 'autonomous orchestrator|delegate|foundation:' .amplifier/context/build-session.md

Primary contributors: Sam Schillace authored 283 of 284 commits; 1 commit by Michael J. Jabbour. 296 commits carry the Co-Authored-By: Amplifier trailer.

Gaps: The ~62h figure is wall-clock and includes a 38.3h gap; continuous coding was ~24h. Compaction counts, sub-agent counts, and token figures were NOT reproducible from these repos and are deliberately not claimed. Test pass rate is not verifiable here (runtime deps not installed); only the static count of 1,454 test functions is verified.