[FEAT][EXP] Adaptive optimizations (CF-831)

[mohammedahmed18]

depends on https://github.com/codeflash-ai/codeflash-internal/pull/2157

[mohammedahmed18]

TODO: instead of checking the tree length, check if there was a refinement before, and only call the adaptive endpoint if so.

the reason is that calling the adaptive on non-refined candidate is a bit of duplication, as the candidate will be refined anyway

so we can end up having these two processes in parallel

optimize (c1) -> repair (c2) -> adaptive (c3)
optimize (c1) -> repair (c2) -> refine (c3) -> adaptive (c4)

resulting in 3 different candidates (2 adaptives, 1 refinement)

it's better to only have
optimize (c1) -> repair (c2) -> refine (c3) -> adaptive (c4)

which will result in 2 candidates (1 refined , 1 adaptive)

[KRRT7]

Code Review: Adaptive Optimizations

Potential Issue: Forest Cycle/Corruption

In CandidateForest.add(), if /adaptive_optimize returns a candidate with a parent_id that doesn't exist yet, a placeholder node is created:

if parent is None:
    parent = CandidateNode(candidate=None)  # placeholder
    self.nodes[pid] = parent

This placeholder has candidate=None, but path_to_root() unconditionally appends node.candidate:

def path_to_root(self) -> list[OptimizedCandidate]:
    path = []
    node: CandidateNode | None = self
    while node:
        path.append(node.candidate)  # Will append None for placeholder
        node = node.parent
    return path[::-1]

Then in call_adaptive_optimize(), the code iterates over this path and accesses c.optimization_id, c.source_code, etc. - which would raise AttributeError if c is None.

Question: What prevents the adaptive optimize endpoint from returning a candidate with a parent_id that hasn't been processed yet? If this can happen, path_to_root() needs to filter out None candidates or the placeholder logic needs rethinking.

[KRRT7]

used Claude to help me review here, testing it for that

[mohammedahmed18]

@KRRT7 in adaptive optimization the new candidate's parent is the last candidate in the tree
so any adaptive optimization we are sure it has a parent

https://github.com/codeflash-ai/codeflash-internal/blob/5d938dd3f9afe86c9d163eef97b4c28ae2ac7388/django/aiservice/adaptive_optimizer/adaptive_optimizer.py#L103-L105

[aseembits93]

@mohammedahmed18 why do we need a new endpoint for adaptive optimization? why can't we modify existing endpoints to take past optimizations into consideration?

[KRRT7]

@aseem then we'll have to keep those optimizations in memory in the backend which would be tricky to implement i guess

[mohammedahmed18]

@aseembits93 @KRRT7 actually I'm using the same optimizer context used in optimize & line-profiler optimize.

https://github.com/codeflash-ai/codeflash-internal/blob/main/django/aiservice/adaptive_optimizer/adaptive_optimizer_context.py

the duplication we have in all 5 endpoints can be refactored but maybe not in a single endpoint imo, this will make it hard for us to implement something specific for a single endpoint.