Skip to content
This repository was archived by the owner on Nov 8, 2026. It is now read-only.

BASELINE.md

Latest commit

 

History

History
46 lines (35 loc) · 2.6 KB

BASELINE.md

File metadata and controls

46 lines (35 loc) · 2.6 KB
Raw

Baseline: what we know about current app-building cost

Written 2026-10-22. Short version: the 12 hours to reach 50 % figure is the owner's experience and is not recoverable from git history. This file records what history does and does not show, so Phase 0 measures instead of guessing.

What git history shows (origin/dev)

App Commits First to last on dev Note
rust/claude-pantheon-rs 40 10-21 20:51 to 10-22 01:46 (~5 h) Preceded by a ~1 h spike in a standalone checkout (10-21 19:55 start, 20:00 backend decision). Feature-rich within ~6 h end to end, with the T3 server as a given backend
rust/mseries-diskspace 1 (+1 housekeeping) single commit 10-21 17:33 Squashed in; the work time is not visible
rust/mseries-updates + -core 1 (+1 housekeeping) single commit 10-21 16:27 Squashed in
mseries/claude-pantheon-native (Vala twin) 1 10-21 20:36 Squashed in
mseries/quick-settings 16 10-12 to 10-21 Plugin work spread over nine days alongside other tasks

Findings:

  1. A large Rust/Granite app can be built fast (the Claude client: ~6 h). So the cost is not uniform; it is high for some apps and some kinds of work. We don't yet know which.
  2. Squashed commits and stacked WIP branches hide real agent time. Commit timestamps measure integration, not effort.
  3. The standalone claude-pantheon-rs and claude-pantheon checkouts have only 10 and 4 commits. They also don't contain the working time.

What would show it

Per-session facts live in the agent session logs (Claude Code transcripts under ~/.claude/projects, Codex and OpenCode logs), not in git. Phase 0 therefore does two things:

  1. Retro-measure two past apps from session transcripts: wall-clock from the first prompt to the first window that opened, to the first state that passes half the acceptance lines, and to done; plus tokens, retries and human interventions. Candidates: Disk Space and Software Updates (both squashed, so history alone says nothing). Record the transcript ids used.
  2. Run specs s01 and s03 fresh with the current process on a frontier agent, using the rubric, to get a controlled "before" that doesn't depend on memory.

Exit: a table here with real numbers and transcript or run ids. Until then, treat every "hours" figure in PLATFORM-PLAN.md as the owner's report, and every "minutes" figure as a proposed target.

Results table (fill in)

Run Spec Mode Model t_first_window t_done Done-bar items met Interventions Notes