# Antigravity transcript tail-boundary proof

Base: 524c757f827a869e42bf9e59ebb8c8cd4f0233d1
Head: bb326cab5953630dd749344ee29008dd3f83ad5f

Environment: Linux web dev app, Chromium preview browser, 1280 x 800 CSS pixels,
2x device scale, light theme. Each commit ran its own real server and fresh database.
Antigravity CLI 1.5.13, Gemini API-key authentication, Gemini 3.8 Flash through the test hub.
No database was copied or modified directly. No tests were used as proof.
The generic ACP replay provider cannot create Antigravity native child records, so this
used the manual two-checkout route.

On each commit:
1. Add Antigravity in Settings > Providers > Add provider. Choose Gemini API key,
   enter a synthetic nonempty key, add the instance, and Connect.
2. Create a scratch git project containing notes.txt with `The marker number is 4242.`
   Add it with the app's project CLI. Start a real thread with:
   `Launch exactly one subagent with role Searcher. Ask Searcher to read notes.txt and report the marker number to you. Wait for Searcher to finish and report its result. Do not read notes.txt yourself.`
   Use provider antigravity, model gemini-3.8-flash, title Transcript boundary proof,
   and the app's `thread start --wait --json` CLI. Both turns completed and reported 4242.
3. Open the parent in the web client, expand Lineage > Previous agents, open Searcher,
   then Subagent work. The unmodified child transcript was readable on both commits.
   Base had six native records; head had eight because its initial find command timed
   out on a permission prompt before it read the file. The synthetic boundary fixtures
   below replace that unrelated variation with identical retained records.
4. Run shape-transcript.py against that child's transcript.jsonl with mode `two`.
   It retains the original opening four records and appends complete steps 4 and 5,
   using an ignored JSON field for byte padding. Those last two lines, including their
   newlines, total exactly 262144 bytes; the preceding byte is a newline.
   Press Refresh. Base shows only step 5; head shows steps 4 and 5.
5. Run the same script with mode `one`. Step 5 alone, with its newline, is exactly
   262144 bytes after a newline. Press Refresh. Base shows no steps; head shows step 5.
6. Additional checks: mode `partial` starts the window inside step 4 and leaves a
   complete step 5; both show only step 5. Mode `undelimited` has no newline in its
   retained window; both show no steps. Every truncated case retains the indicator.

The two-record retained windows were byte-identical across commits (SHA-256
63f79460a0a5f1c701a381f8089fb94f863f2fb5f3edd9cf56cd0e49c2cfcedf).
The script also makes the one-record windows byte-identical regardless of opening data.
The model-authored launch prompt differs slightly between runs; it is outside the changed reader.

Capture procedure: screenshots were taken after matching dialog captures more than
500 ms apart. Images are cropped to the complete dialog. Recordings show the same
sequence: two-record fixture, Refresh, replace with one-record fixture, Refresh.
Videos use the same center crop to omit unrelated environment labels. No app DOM or
rendered content was changed for a capture. Every final image and recording was visually
reviewed. Browser interactions used the shared preview browser; selectors were inspected
from the actual DOM. Synthetic fixtures and this script were not committed.

Not independently verified: 4000-character step text and 1000-character tool-input limits;
padding uses an ignored field. The internal no-newline decoder short circuit has no visible
before/after distinction, although its empty UI result was exercised on both commits.
