| name | antigravity-handoff |
| description | Pick up context from a Gemini Antigravity session to continue its work. Triggers on: hit quota/limit mid-flight, continue Antigravity work, resume from Antigravity, pick up the context, a supplied brain directory, ~/.gemini/antigravity/brain/, transcript_full.jsonl, an Antigravity handoff, "finish what the other agent started". |
Antigravity Handoff — Resume Work From a Gemini Antigravity Session
Antigravity runs a session in a brain directory and does the actual repo work in a worktree. Your job on handoff: recover its understanding cheaply, verify it against ground truth, and continue.
Governing principle: the repo and the brain artifacts ARE Antigravity's externalized memory. Read the distillations, don't replay the episode — then cross-check them against code and git, because the agent drifts (it leaves work uncommitted, introduces broken links, and its prose goes stale). Mismatches between prose, code, and git are the highest-value findings, not the agreements.
Check the repo first — the handoff may already be archived
Antigravity's own session-handoff skill archives the brain's walkthrough.md / implementation_plan.md / task.md into the repo at docs/sessions/YYYYMMDD/HHMMSS-title.md at session end (its record step; start replays the newest one). So the distilled handoff often already lives in the worktree. Read the newest docs/sessions/** file before parsing the raw brain transcript — it's the curated version of exactly what you're reconstructing. Fall back to the brain directory / transcript_full.jsonl when sessions/ is missing, stale, or the session was cut off before record ran (a mid-flight quota cutoff is common — so always still check git status for uncommitted work).
The brain directory (what's inside)
~/.gemini/antigravity/brain/<uuid>/
.system_generated/logs/transcript_full.jsonl — full reasoning trace, one step per line. Large (100KB+). Never read linearly — grep/sed by step_index.- Root artifacts:
implementation_plan.md, task.md, walkthrough.md — intent, current-state, and how-to deliverables.
scratch/ — throwaway analysis scripts. evaluation_games/, domain outputs (*_analysis.md).- The brain references its repo via
workspaceUris → a worktree under ~/.gemini/antigravity/worktrees/.
transcript_full.jsonl schema (per line)
{step_index, source, type, status, created_at, content, thinking, tool_calls}
source: USER_EXPLICIT | MODEL | SYSTEMtype: USER_INPUT (content wrapped in <USER_REQUEST>…), PLANNER_RESPONSE (has thinking + tool_calls), VIEW_FILE/RUN_COMMAND/GREP_SEARCH/CODE_ACTION (tool results), CHECKPOINT, ERROR_MESSAGE.CHECKPOINT steps are gold — they summarize truncated context: outstanding user requests, a session summary, the artifact list, and spawned subagents.- Finding the cut-off: the tail often ends in repeated
ERROR_MESSAGE (error_code:429) with empty MODEL steps. The last real intent is the last USER_INPUT; the last landed change is the last CODE_ACTION. Work often lands but is never committed — always check git status.
Useful probes (don't dump the whole file):
wc -l transcript_full.jsonl; grep -n '"type":"CHECKPOINT"'; grep -n '"type":"USER_INPUT"' to map the spine.grep -n '<step_index>' … then sed -n 'A,Bp' to read a window around a point of interest.
Read strategy — cheap spine first, drill on demand
| Tier | Read | How |
|---|
| 0 — spine (do unconditionally) | newest docs/sessions/** file (archived handoff, above); brain task.md, implementation_plan.md; repo docs/**/master_*history*.md (pivot log); git log --oneline -30; git status | Read fully — cheap, recovers ~80% of orientation |
| 1 — procedural | the project's own agent skills (.agents/skills/, .claude/skills/) | Skim |
| 2 — rationale (on demand) | docs/qna/*.md or similar design docs | Index by title; read only the 2–3 touching the current task |
| 3 — ground truth (verify, don't absorb) | the exact functions the docs/skills point at | Read the referenced function, not the whole file |
| 4 — episodic fallback | transcript_full.jsonl | grep only, for a specific question Tiers 0–3 couldn't answer |
Verification discipline (this is what beats Antigravity)
For anything load-bearing, triangulate three independent sources: prose (docs) vs. code (functions) vs. timeline (git log + disk state). Antigravity accumulated context but couldn't cheaply self-check. When they disagree, the disagreement is the finding — surface it before continuing.
Stopping rule (keeps context lean)
Stop reading once you can answer without guessing:
- What are we ultimately building? 2. What state/phase are we in now? 3. What's the immediate next action? 4. What's been tried and rejected (don't repeat it)? 5. How could the next step break?
Pull context against a live question; never hoard it in advance. If a question survives Tiers 0–3, then grep the transcript.
