362 sessions with Pi: the agent failed at least once in 120, and 193 finished without a human stepping in.
Takeaway: Next time I'd ask directly for the root cause investigation across the transcript rendering pipeline instead of just reporting the visual bug, since the fix ended up touching several files (Vue component, transcript utils, bridge types).
Takeaway: Next time I'd attach the screenshot and explicitly describe what looks wrong (e.g.
Takeaway: Next time I'd open with the concrete ask up front — 'add a way to open any folder as a workspace, not just ones already scanned in .pi/agent/sessions' — instead of first describing the limitation and only later asking for the fix, since that cost a whole extra round of re-exploration.
Takeaway: I'd start by just asking for a one-line summary plus the key entry points instead of letting the agent read every single file, since most of the answer came from README.md and a handful of core files.
Takeaway: Next time I'd point straight at extensions/pi-traceline/index.ts and docs/design-language.md instead of letting it grep the whole repo first, since that's where the emphasis logic and tone table actually lived.
Takeaway: Next time I'd put the design contract itself up front in the first prompt — asking for a written design-language doc before any code — since that's what let the later passes move fast once agreed.
Takeaway: I'd keep this prompt as-is next time since it's a simple smoke test of file reading, writing, and subagent delegation.
Takeaway: I'd open with the concrete repro and screenshot up front rather than asking 'why' first — the JSONL forensics dig was the right instinct but could've started immediately.
Takeaway: I didn't actually give the agent a real task here, so there's not much to judge yet.
Takeaway: I'd start by just naming exactly which extension behaviors I want tested up front (read, write/delete, subagent, scratchpad) rather than spreading it across three prompts.
Takeaway: I'd start the prompt the same way next time, since it worked well: give the exact broken example and be explicit that I want diagnosis, not a fix, so the agent doesn't overreach into UI/rendering solutions.
Takeaway: I didn't actually give a task here, so there's nothing to evaluate yet.
Takeaway: I didn't actually give the agent a task in this session, just said hi.
Takeaway: Next time I'd ask the tool-search question up front instead of first asking 'did we make a mistake' — the real question was always 'is this duplication necessary for a tool-search conversion', and getting there took a long detour through git archaeology and PR diffing.
Takeaway: I'd start the prompt by naming both repos and the exact question (how do the token estimates work, and are we doing it right) instead of exploring cc-inv first and only later remembering to compare against Codex and our own code.
Takeaway: I'd start next time by directly asking to pull and diff the actual CLI source for both agents instead of starting with docs, since that's what actually settled the questions (unpacking the npm tarballs and running `amp tools show`).
Takeaway: I'd open next time by asking directly for a script that captures raw provider payloads alongside token usage, since that's what actually let us settle the hidden-content question.
Takeaway: I just wanted a quick sanity check on naming, and honestly I could have pointed the agent straight at the two extension READMEs instead of letting it crawl the whole repo tree first.
Takeaway: Next time I'd ask the agent to also check whether the new assertion helper actually tightens the bound relative to the scheduling latency, not just whether it removes the sleep.