Repository navigation
Replies: 10 comments
|
Some real numbers from trying it, plus a resume constraint that affects how T3 should expose it. Measured savings. Same task, same repo snapshot, two interactive
That's about half the cached input tokens and a third less time. The context didn't shrink much: the model saw roughly the same volume of tool output in both runs. The saving comes from fewer round trips. In the Code Mode run, 24 Resume is locked to the mode a session started in (checked on 2.1.323):
So once the Agent SDK supports it, the setting should belong to the thread: chosen when the thread starts and fixed after its first turn, since changing it would need a fresh native session. An environment variable on the Claude provider instance would switch every new thread at once, and turning it off later would leave existing Code Mode threads unable to resume. I'm happy to put up a PR for a per-thread option once the SDK can run Code Mode. |
|
+1, with numbers from one heavy T3 user: 30 days of Claude threads in T3 (about 1,000 sessions and 8,200 prompts), counted from session transcripts. 1. Mostly one tool call per step.
That's the round-trip pattern Code Mode removes. 2. Subagents make more requests than main threads.
3. Multi-step skills drive the longest chains.
4. Waiting is my most-used T3 tool.
|
|
+1 |
|
+1 |
|
+1 |
|
+1 |
|
+1 |
|
+1 |
|
Note 🤖 Claude Opus 5.5 responding on behalf of Jonathanm10 The Agent SDK route already exists in the Claude Code CLI. What blocks T3 is its What the CLI does (Claude Code 2.1.330, read from the native binary's strings):
Reproduction: claude -p --settings '{"env":{"CLAUDE_CODE_EXPERIMENTAL_CODE_MODE":"1"}}' \
--output-format stream-json --verbose "Reply with just OK"
On the T3 side:
Possible fixes: (1) T3 sends an SDK entrypoint, at least for threads that opt into Code Mode. (2) Anthropic adds |
|
+1, and confirming the findings in the previous comment from inside a T3 Claude thread, with a few additions. (Claude Code 2.1.330, native build, Linux, Opus 5.5, Anthropic API.) Confirmed in a live T3 thread
How the opt-in is read (from the binary's strings)
Other refusal reasons T3 may want to surface (from the CLI's refusal-text table)
Suggestion: a per-thread Code Mode toggle, fixed at thread start because of the resume lock described above, that:
Workaround until then: a terminal Investigated with Claude Code. |
Uh oh!
There was an error while loading. Please reload this page.
Claude Code 2.1.323 added Code Mode (experimental): Claude writes JavaScript that calls its tools, so one step can run several calls in parallel and filter results before Claude reads them. Docs: https://code.claude.com/docs/en/code-mode
It isn't available in T3 Code. The docs say it only works in an interactive
claudesession in a terminal, not with-p, the Agent SDK, the VS Code panel, the Desktop app or cloud sessions. T3's Claude threads run through the Agent SDK, so settingCLAUDE_CODE_EXPERIMENTAL_CODE_MODE=1doesn't turn it on there.I'd like to use it from T3. Two possible routes:
All reactions