Upgrade pi to 0.87.1 for Claude 5.5 and GPT-6 - #556
dcartertwo wants to merge 3 commits into
Conversation
Upgrade pi to 0.87.1 and adapt the shared agent loop to its new system-message and turn-completion behavior.
github-actions
Bot
added
kernel
|
LGTM! |
Preview:
|
Eval runs comparisonBaseline No cohort is comparable. Not compared:
|
Eval runs reviewMeasured. No cohort is comparable because all three eval definitions changed, so no measured deltas can be attributed. |
|
LGTM! |
| "claude-haiku-4-5": {name: "Claude Haiku 4.5", contextWindow: 200000}, | ||
| }, | ||
| "openai": { | ||
| // pi's GPT-6 catalog advertises a 272K window; reserve the 128K response cap when sizing |
There was a problem hiding this comment.
Why does it use 272k window instead of full million context with compactionInputBudget of 272k like gpt 5.6-sol?
There was a problem hiding this comment.
clanker said that was the limit. i didn't verify
There was a problem hiding this comment.
All GPT-6 models have a 1,050,000 context limit.
272,000 is the suggested compaction trigger -- that is exactly what compactionInputBudget is for, so we should use that.
(I do think this whole thing needs to be revised a bit but maybe not in this PR.)
There was a problem hiding this comment.
Good catch—updated all three GPT-6 models to a 1,050,000-token context window with a 272,000-token preferred compaction budget (and the existing 128,000-token output cap) in 391f89c. Updated the model and compaction tests and corrected the PR description.
Eval runs reviewMeasured. No cohort is comparable because all three eval definitions changed, so no measured deltas can be attributed. |
|
LGTM! |
![devin-ai-integration[bot]](https://avatars.githubusercontent.com/in/811515?s=60&v=4){kind=small}
| // pi's GPT-6 catalog reports a 272K window, but these models support 1.05M. Use 272K as the | ||
| // preferred compaction budget, not as the hard context limit. |
There was a problem hiding this comment.
🔍 Conflicting GPT-6 context limits
contextWindow overrides pi’s 272K catalog value with 1.05M. The PR description still defines 272K as the hard window, so provider confirmation is needed.
Was this helpful? React with 👍 or 👎 to provide feedback.
Eval runs reviewMeasured. No cohort is comparable because all three eval definitions changed, so no measured deltas can be attributed. |
4 participants
Upgrade pi to 0.87.1 to add Claude Opus 5.5, Claude Fable 5.1, and GPT-6 Astra, Sol, and Luna to Workshop’s suggested models, with catalog-backed cost estimates and the provider settings needed to run them correctly. Keep GPT-6’s 1.05M context window and 128K output limit, using 272K as the preferred compaction budget. Quick requests use low effort on Claude models that require adaptive thinking.
pi 0.87 now expects agent instructions and tool descriptions in the conversation, and asks Workshop whether to continue before signaling the end of a turn. We updated the shared agent loop to supply that context and capture pending approvals in time to save them with the turn. This preserves guidance for every model and lets agents pause for approval or a new connection with their work recorded.