plan mode: add TL;DR checkpoint and client behavior note #10195
Conversation
|
I have read the CLA Document and I hereby sign the CLA Nick Baumann seems not to be a GitHub user. You need a GitHub account to be able to sign the CLA. If you have already a GitHub account, please add the email address used for this commit to your account. |
|
|
||
| Begin by grounding yourself in the actual environment. Eliminate unknowns in the prompt by discovering facts, not by asking the user. Resolve all questions that can be answered through exploration or inspection. Identify missing or ambiguous details only if they cannot be derived from the environment. Silent exploration between turns is allowed and encouraged. | ||
|
|
||
| Before asking the user any question, perform at least one targeted non-mutating exploration pass (for example: search relevant files, inspect likely entrypoints/configs, confirm current implementation shape), unless no local environment/repo is available. |
There was a problem hiding this comment.
good call on removing some things -- testing tonight it feels this was doing a lot of the heavy lifting in preventing premature plans
| * Keep asking until you can clearly state: goal + success criteria, audience, in/out of scope, constraints, current state, and the key preferences/tradeoffs. | ||
| * Bias toward questions over guessing: if any high-impact ambiguity remains, do NOT plan yet—ask. | ||
|
|
||
| ## TL;DR checkpoint (required, before PHASE 3) |
There was a problem hiding this comment.
let's remove this for now. and focus on final output
| * Status updates must not include questions or plan content. | ||
| * Internal tool/repo exploration is allowed privately before A, B, or C. | ||
|
|
||
| Client behavior note (important): when a turn completes with normal assistant text (anything other than `request_user_input`), that text is treated as plan content by the client. During clarification, prefer `request_user_input` for user-facing back-and-forth; keep status updates brief and non-plan. |
There was a problem hiding this comment.
That would confuse the model for preambles, let's remove it for now.
| * not be answerable by non-mutating commands. | ||
|
|
||
| Use the `request_user_input` tool only for decisions that materially change the plan, for confirming important assumptions, or for information that cannot be discovered via non-mutating exploration. | ||
| Use the `request_user_input` tool only for decisions that materially change the plan, for confirming important assumptions, or for information that cannot be discovered via non-mutating exploration. Always set `isOther: true` so freeform input is available. |
There was a problem hiding this comment.
let's keep the freedom to the model
| ## Finalization rule | ||
|
|
||
| Only output the final plan when it is decision complete and leaves no decisions to the implementer. | ||
| Only output the final plan when it is decision complete, approved via the TL;DR checkpoint, and leaves no decisions to the implementer. |
There was a problem hiding this comment.
let's not have checkpointing and focus on output and stopping the premature plans
There was a problem hiding this comment.
having it at the top of the streamed plan accomplishes a similar goal here
| The final plan must be plan-only and include: | ||
|
|
||
| * A clear title | ||
| * A TL;DR section (3–5 bullets) consistent with the approved TL;DR checkpoint (unless the user changed it) |
There was a problem hiding this comment.
I feel we can even remove * Exact file paths to change and
make this less verbose: * Exact function, method, type, and variable names and signatures (only important signatures or structs?)
Summary
What changed
-isOther: true, with explicit guidance that “None of the above” is the edit path in the current UI.Why
Testing