@@ -19,8 +19,15 @@ mkdir -p "$HOME/.local/bin"
 ln -sfn "$go_root/bin/go" "$HOME/.local/bin/go"
 ln -sfn "$go_root/bin/gofmt" "$HOME/.local/bin/gofmt"
 
-if ! "$go_root/bin/go" version -m "$(command -v bd 2>/dev/null || echo /dev/null)" 2>/dev/null | grep -q "${bd_commit:0:12}"; then
+if ! "$go_root/bin/go" version -m "$HOME/.local/bin/bd" 2>/dev/null | grep -q "${bd_commit:0:12}"; then
   CGO_ENABLED=1 GOFLAGS=-tags=gms_pure_go GOBIN="$HOME/.local/bin" \
     "$go_root/bin/go" install "github.com/steveyegge/beads/cmd/bd@$bd_commit"
   ln -sfn "$HOME/.local/bin/bd" "$HOME/.local/bin/beads"
 fi
+
+repo_root="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")/.." && pwd)"
+git_common_dir="$(git -C "$repo_root" rev-parse --path-format=absolute --git-common-dir)"
+beads_repo_root="$(dirname -- "$git_common_dir")"
+if [[ ! -d "$beads_repo_root/.beads/embeddeddolt/beads/.dolt" ]]; then
+  "$HOME/.local/bin/bd" --directory "$beads_repo_root" bootstrap --yes
+fi
@@ -0,0 +1,74 @@
+---
+name: analyzing-orb-sizing
+description: Retrospectively assesses whether an Amp orb was under-sized, over-sized, or appropriately sized from thread tool output and measured workloads. Use when reviewing orb size, resource pressure, OOM failures, build performance, or orb cost efficiency.
+compatibility: Requires Python 3.9+ and the Amp CLI when analyzing a thread ID.
+argument-hint: <thread-id-or-export.json> --size <a0.tiny|a0.small|a0.medium|a0.large>
+---
+
+# Analyzing Orb Sizing
+
+Assess orb sizing from evidence, not repository type alone. Run the bundled analyzer first, then use `read_thread` when semantic context is needed to decide whether observed workloads were representative.
+
+## Run the analyzer
+
+From the repository root:
+
+```bash
+python3 .agents/skills/analyzing-orb-sizing/scripts/analyze.py \
+  T-xxxxxxxx-xxxx-xxxx-xxxx-xxxxxxxxxxxx --size a0.small
+```
+
+It also accepts an existing export, which is useful offline:
+
+```bash
+amp threads export T-xxxxxxxx-xxxx-xxxx-xxxx-xxxxxxxxxxxx > /tmp/thread.json
+python3 .agents/skills/analyzing-orb-sizing/scripts/analyze.py \
+  /tmp/thread.json --size a0.small --json
+```
+
+Always pass the size used by that thread. Project defaults are prospective and may not match an existing orb.
+
+## Interpret the result
+
+- **under-sized**: hard resource-pressure evidence such as OOM, exit 137, allocation failure, or measured usage that reaches the orb limit. Move up one size, then repeat the representative workload.
+- **over-sized**: representative measured workloads fit comfortably on the next smaller size with CPU and memory headroom. Move down one size and verify there.
+- **right-sized**: representative measurements use meaningful capacity without pressure and do not safely fit the next smaller size.
+- **insufficient-evidence**: the transcript lacks trustworthy resource measurements or hard pressure signals. Do not interpret an uneventful thread as proof of over-sizing.
+
+The analyzer only treats shell/tool results as machine evidence. User and assistant prose can describe the task but must not be counted as telemetry.
+
+## Add evidence to a future thread
+
+Run the slowest representative build or test under GNU `time` inside the orb:
+
+```bash
+/usr/bin/time -v cargo test
+/usr/bin/time -v ./gradlew test
+```
+
+For parallel builds, also record the orb identity and capacity:
+
+```bash
+printf 'orb-capacity cpus=%s memory_kb=%s\n' "$(nproc)" "$(awk '/MemTotal/ {print $2}' /proc/meminfo)"
+```
+
+Use a clean and an incremental build when both matter. One trivial command, setup/install work, model thinking time, network waits, and idle thread duration are not representative sizing measurements.
+
+## Apply judgment after the script
+
+1. Use `read_thread` to identify what the measured command did and whether it represents normal project work.
+2. Separate compute delay from dependency downloads, network services, lock contention, and test sleeps.
+3. Prefer multiple representative samples. A hard OOM is decisive; an over-sizing recommendation needs measured headroom.
+4. Mention confidence, evidence, and missing evidence in the final recommendation.
+5. Disk is 40 GB for every documented size, so changing size does not solve disk pressure.
+
+## Size reference
+
+| Size | CPUs | Memory | Hourly price |
+|---|---:|---:|---:|
+| `a0.tiny` | 1 | 2 GB | $0.10 |
+| `a0.small` | 2 | 4 GB | $0.21 |
+| `a0.medium` | 8 | 16 GB | $0.83 |
+| `a0.large` | 16 | 32 GB | $1.66 |
+
+Treat prices as a dated reference and verify current Amp pricing before making a cost forecast.