You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Antigravity skill loads show as raw Read calls on T3's hashed per-instance profile path #40639
A turn that loads several skills shows a stack of these rows, each with a 64-character hash in the middle (the SHA-256 of the instance ID, from resolveAntigravityInstanceDirectories), so you can't tell at a glance which skills the agent used. Expanding a row shows only the tool input JSON ({"toolName":"Read","input":{"Path":"…","path":"…"}}), with the same path repeated.
For Claude threads, the Skill tool call shows as Skill: <name>.
Diagnosis
Antigravity has no skill tool. The model loads a skill by reading <profile>/config/skills/<name>/SKILL.md with its normal Read tool. Those profile entries are symlinks T3 creates to the user's skill folders.
packages/shared/src/toolActivity.ts gives skills their own title only for Claude (claudeSkillInvocation, which needs toolName === "Skill", and becomes Skill: ${name}). A read goes through formatReadToolLabel, which prints the full path.
So a read of a SKILL.md under the provider's own skills directory (or any known skill root) gets no skill title. Showing it as Skill: <name>, or at least replacing the profile root with something short, would make these turns readable.
Unchanged on main as of 2026-11-22.
Steps to reproduce
Set up an Antigravity provider instance with a few user skills linked into its profile.
Start a thread with a Gemini model (seen with gemini-4-neon) and ask for something that brings in several skills, e.g. "rebuild this project in line with our skills".
Look at the timeline: each skill load is a Read …/providers/antigravity/<hash>/config/skills/<name>/SKILL.md row, and expanding it shows only the input JSON.
Version
0.1.12-nightly.20261121.5064 (desktop app and local server)
Environment
Linux x64 (CachyOS, kernel 7.3.3), T3 Code Desktop nightly with a local server, Antigravity provider, model gemini-4-neon.
Evidence
One turn in a single thread loaded typescript-stack, slop-indicators, desktop, effect, setup-docs, and testing. Each shows up as a separate Read <profile path>/config/skills/<name>/SKILL.md row with the input JSON as its only detail. Screenshot attached below.
I tried to reproduce this on the current build and could not.
Commit 83edd5e (origin/main). Ran T3CODE_DEV_SEED=0 GOOGLE_GEMINI_BASE_URL=<hub> HOME=<scratch home> vp run dev in a fresh worktree, so the server read a sandboxed ~/.gemini/config/skills/pineapple-check/SKILL.md. In the web client: Settings > Providers > Add provider > Antigravity, sign-in method 'Gemini API key' with a dummy key, then Connect. The status became 'Authenticated · Gemini API key' (runtime v1.5.14). The server created the instance profile .t3/userdata/providers/antigravity/ac0a3dfd6dddb20962cecff6ee5fe65e19d3923be20e52c5ab52ff877f7e4c32/ (the same hash as the report), with config/skills linked to the user skill folder. Started a thread asking the model to read the pineapple-check SKILL.md, using Gemini 4 Neon, then Gemini 4 Hydrogen, then Gemini 3.8 Flash. Each turn failed with hub error 429 RESOURCE_EXHAUSTED ('All credentials for model ... are cooling down', reset in about 96h), so no live skill-read row could be produced. Instead I wrote a unit test on the real projection path: a completed dynamic_tool item (the type ACP 'read' calls become, per canonicalItemTypeFromAcpToolKind) with toolName 'Read' and input {Path, path} pointing at /config/skills/effect/SKILL.md, passed through deriveTimelineEntriesFromVisibleTurnItems and workEntryDisplayLabel. On main it returns 'Read /home/user/.t3/userdata/providers/antigravity/ac0a3dfd…c32/config/skills/effect/SKILL.md' instead of 'Skill: effect'.
Closing as not planned. If it still happens for you, reply with the exact steps and the version you are on and it will be looked at again.
Activity
juliusmarminge commentedon Nov 22, 2026
I tried to reproduce this on the current build and could not.
Commit 83edd5e (origin/main). Ran
T3CODE_DEV_SEED=0 GOOGLE_GEMINI_BASE_URL=<hub> HOME=<scratch home> vp run devin a fresh worktree, so the server read a sandboxed ~/.gemini/config/skills/pineapple-check/SKILL.md. In the web client: Settings > Providers > Add provider > Antigravity, sign-in method 'Gemini API key' with a dummy key, then Connect. The status became 'Authenticated · Gemini API key' (runtime v1.5.14). The server created the instance profile .t3/userdata/providers/antigravity/ac0a3dfd6dddb20962cecff6ee5fe65e19d3923be20e52c5ab52ff877f7e4c32/ (the same hash as the report), with config/skills linked to the user skill folder. Started a thread asking the model to read the pineapple-check SKILL.md, using Gemini 4 Neon, then Gemini 4 Hydrogen, then Gemini 3.8 Flash. Each turn failed with hub error 429 RESOURCE_EXHAUSTED ('All credentials for model ... are cooling down', reset in about 96h), so no live skill-read row could be produced. Instead I wrote a unit test on the real projection path: a completed dynamic_tool item (the type ACP 'read' calls become, per canonicalItemTypeFromAcpToolKind) with toolName 'Read' and input {Path, path} pointing at /config/skills/effect/SKILL.md, passed through deriveTimelineEntriesFromVisibleTurnItems and workEntryDisplayLabel. On main it returns 'Read /home/user/.t3/userdata/providers/antigravity/ac0a3dfd…c32/config/skills/effect/SKILL.md' instead of 'Skill: effect'.Closing as not planned. If it still happens for you, reply with the exact steps and the version you are on and it will be looked at again.