Explain conda-forge bazel-feedstock repo structure
Takeaway: I just wanted a quick orientation to a repo I hadn't touched before, and asking 'what is this repository?' worked fine to get a fast, accurate overview.
Agent · 362 sessions
Real Pi sessions
362 sessions with Pi: the agent failed at least once in 120, and 193 finished without a human stepping in.
Fix conda-forge bazel-feedstock linux-aarch64 build failure
Takeaway: Next time I'd start by dumping the installed packages in the build prefix earlier, since the duplicate systemlib packages were the actual root cause and cost a lot of back-and-forth with conda/python/pixi path issues.
Package torch-c-dlpack-ext for conda-forge via rattler-build
Takeaway: Next time I'd give the agent the newer build-helper script location upfront, since a lot of the back-and-forth was chasing down _build_optional_torch_c_dlpack.py and the right compiler/build-step config through trial and error with rattler-build.
Trace wheel build source in tvm-ffi CI workflow
Takeaway: Next time I'd start by asking it to trace the checkout step and build script together instead of one file at a time, since the answer hinged on both.
Build a rattler-build recipe for runai-model-streamer with Bazel C++ core
Takeaway: Next time I'd tell the agent upfront that the target build host is linux-aarch64 (not macOS) so it skips the Docker/cross-compile detour and gets straight to patching the WORKSPACE.
Fix rattler-build npm/pnpm error in conda-forge feedstock (linux-aarch64)
Takeaway: Next time I'd tell the agent upfront to inspect the actual extracted source tree before writing a patch, since guessing at line numbers from a stale copy wasted two build cycles.
Fix libxmlsec1 conda-forge build failures across platforms
Takeaway: Next time I'd give the agent push rights or a way to open the PR directly, since we hit a 403 at the end after all the fixing was done.
Set macOS-26 GitHub Actions runner in conda-forge feedstock
Takeaway: I'd tell the agent upfront to check conda-smithy's GITHUB_ACTIONS_RUNS_ON mapping directly instead of spending several web searches first, since the source code fetch is what actually gave the answer.
Write a python CLI to upload pi agent traces to Hugging Face
Takeaway: I'd specify pathlib from the start instead of asking for it as a follow-up correction, since that's just a style preference I already have.
Create conda-forge recipe for apple/container Swift package
Takeaway: Next time I'd tell it up front that macOS ships BSD install, not GNU coreutils, since that's the kind of platform detail it can't discover just by reading the recipe skill and docs.
Fix go-licenses build failure in opentofu conda-forge feedstock
Takeaway: I'd start the prompt by pointing directly at the go-licenses skill reference since that's what actually solved it — the fix was a known workaround pattern for packages with undetected licenses.
Fix conda-forge go-licenses failure in process-compose feedstock
Takeaway: I'd start the prompt by pointing straight at the conda-forge skill's go-licenses reference since that's exactly where the fix came from — it saved a lot of guesswork.
Fix rattler-build failure for process-compose conda-forge feedstock (linux-aarch64)
Takeaway: Next time I'd point straight at the go-licenses error and check the skill's reference doc first, since that's exactly the known failure mode for chroma/v2's COPYING file.
Fix conda-forge process-compose feedstock aarch64 build failure
Takeaway: Next time I'd start by telling the agent straight away that the failure is likely a go-licenses classification issue for a Go dependency, since that's a known conda-forge pattern.
Fix conda-forge process-compose aarch64 build failing on go-licenses
Takeaway: Next time I'd tell the agent upfront that go-licenses chokes on dependencies with non-standard COPYING files mixing MIT and OFL-1.1, since that's exactly what killed this build.
Fix pnpm check:web type errors from mismatched mermaid API
Takeaway: Next time I'd point straight at the file using beautiful-mermaid instead of just saying 'fix pnpm check:web' — the agent spent a lot of turns reading through MarkdownRenderer.svelte before it found the real issue, that the code was written for classic mermaid.js but imported a completely different library.
Replace Comark with marked and trim Shiki bundle in Svelte web app
Takeaway: Next time I'd start the first prompt by asking for a full audit of what each bundling layer (parser, languages, themes) actually pulls in, instead of chasing library swaps first.
Revert comark markdown renderer back to marked in Svelte package
Takeaway: Next time I'd ask upfront for the revert plus a rebuild/lint check in one go, since resolving the three-way conflicts in package.json, the lockfile and the component took most of the turns.
Clean up Vite/tsconfig configs across pi-web monorepo packages
Takeaway: Next time I'd start by asking the agent to scan all package configs at once instead of going file by file, since the bin/bridge/svelte tsconfigs and vite configs kept drifting out of sync with each other.
Fix @pi-web/bridge subpath exports for svelte package
Takeaway: Next time I'd start the prompt by pointing directly at the package.json exports field since that's where the real fix ended up living, rather than letting the agent explore the whole monorepo structure first.