peltier.io/CLAUDE.md

hotpath — profile-guided optimization agent

Working name. Rename freely; nothing depends on it.

Mission

An agent that profiles real binaries and services, isolates hot paths, proposes optimizations, and accepts a change only when (a) behavioral-equivalence gates pass and (b) the benchmark improvement is statistically significant on trusted infrastructure. Output: verified performance deltas translated into ROI reports nobody can dispute.

Full engineering spec: SPEC.md. Read it before starting any phase.

Non-negotiables (re-read every session)

1. No unverified performance claims. A change without passing equivalence gates AND a significant bench delta does not exist. Discard it. Log it.
2. Measure before optimizing. No patch is proposed until profile data identifies the hot path. Never trust a single benchmark run.
3. Cheap wins first, always. Build flags → LTO → PGO → allocator swap → then code changes. Check the ledger before re-attempting any class of change on a target.
4. Equivalence gates are hard gates. Changes touching floating-point ordering, concurrency primitives, or anything UB-adjacent are NEVER auto-accepted — verdict is needs-human-review.
5. Every attempt goes in the ledger, including failures and rejections. Failures are future training data.
6. The agent never modifies the trust layer. crates/, config/, corpora/, and upstream test suites are read-only to the agent. Patches may touch only allowlisted paths under targets/<name>/. See SPEC.md §10.
7. Sandbox everything. Target code runs only inside no-network containers. Never on the host.
8. Statistical bar: accept only if the bootstrap 95% CI lower bound of speedup ≥ threshold in config/accept.toml (default 2%), from interleaved A/B runs on the same pinned hardware.

Architecture map

• crates/bench-runner — trust layer: containerized, CPU-pinned, interleaved A/B benchmarking with bootstrap confidence intervals and A/A self-tests
• crates/diff-test — behavioral equivalence: upstream test suite + golden I/O replay + differential fuzzing of changed functions + sanitizers
• crates/ledger — append-only SQLite record of every attempt: hypothesis, patch, gate results, bench deltas, verdict, cost
• crates/report — ROI generator: cores saved × $/core-hr, latency percentile deltas, CIs, workload caveats printed on every number
• agent/ — Claude Agent SDK (Python) loop, prompts, tool definitions
• targets/ — vendored OSS targets for case studies (permissive licenses only)
• playbook/ — optimization classes, ordered, with preconditions and known risks
• results/ — generated reports and flamegraphs

Current phase

Phase 3 COMPLETE — all SPEC §5 exit criteria met. The trust layer went language-agnostic: build isolation moved from a cargo-specific CARGO_TARGET_DIR injection to an explicit {out} placeholder + optional sanitizer/sanitizer_binary templates (crates/diff-test/src/target.rs), proven zero-regression on tokei before any C target. A C target joined the roster — cJSON v1.7.19 (MIT): trust-layer harness driver compiled with the patched cJSON.c, hermetic Unity/CTest gate (19 tests), ASan+UBSan sanitizers, deterministic pinned corpus, own A/A calibration (0/20 FP, 20/20 injection), CI-gated. First verified cross-language win: phase3-cjson-002 (class 5+3, stack-buffer number parse + strtod round-trip) +8.85%, 95% CI [+7.5%, +10.0%] — banked, byte-identical golden replay, independent ASan+UBSan clean, 10k-input diff-fuzz (scripts/diff-fuzz-cjson.py) 0 divergences. Four public case studies in results/phase3/ (tokei class-5 wins; build flags + allocators; the comrak caught false-accept; the cJSON cross-language proof). Six playbook classes exercised (1–6); wins on classes 1/2(review)/3/5 across Rust and C. coz wired (-DHOTPATH_COZ, just coz, scripts/coz-*); its apt runtime aborts in init on this container's glibc, so callgrind cache-sim is the working profiler (documented). Zero regressions shipped. Next: Phase 4 (load replay, latency mode under replayed traffic, mechanical ROI reports — SPEC.md §5; crates/report has the ROI math, needs the generator).

Commands

Fill these in as the harness is built; keep them current:

just profile <target>       # perf record + flamegraph for a target's benchmark workload
just bench <target> <ref>   # interleaved A/B: pristine baseline vs working tree
just gates <target>         # test suite + golden replay + diff-fuzz + sanitizers
just verdict <target>       # runs gates + bench, writes ledger row, prints verdict
just report <run-id>        # ROI report from ledger
just aa <target>            # A/A calibration run (must show null result)

Built so far (command-level, pre-target-integration):

just build / test / lint    # trust-layer workspace: cargo build/test/clippy+fmt
just aa [cmd]               # A/A self-test of bench-runner on a shell command
just compare <a> <b>        # interleaved A/B of two shell commands, bootstrap CI
just gates <target>         # corpus pin + upstream tests + golden replay
just pin-check <target>     # verify corpus MANIFEST.sha256
just calibrate <cmd> <out>  # automated A/A + regression-injection calibration
just verdict <t> <bin> ...  # gates + bench vs pristine-rebuilt baseline + ledger row
cargo run -p harnessd       # agent IPC daemon: six-tool JSON surface (SPEC §3.5)

Definition of done — one optimization attempt

• Hypothesis logged in ledger before patching
• Patch touches only allowlisted paths under targets/<name>/
• Upstream test suite green
• Golden replay byte-identical (or within the target's explicit FP tolerance policy)
• Differential fuzz on changed functions: no divergence (60s or 10k iterations minimum)
• ASan + UBSan clean; TSan if the patch touches anything threaded
• Interleaved A/B bench passes the CI-lower-bound threshold
• Ledger row written with verdict and full evidence
• If FP-ordering / concurrency / UB-adjacent: verdict needs-human-review, never auto-accept

An attempt that fails any gate is a valid, complete outcome — write the ledger row and move on. Do not iterate on a rejected patch more than twice without a new hypothesis.

Conventions

• Rust workspace (edition 2021+) for all harness crates; just for task running
• Python 3.11+ for agent/; Claude Agent SDK — verify current package name and API at https://docs.claude.com/en/docs/claude-code/overview before scaffolding
• Every number shown to a human carries its confidence interval and workload description. No naked percentages, ever.
• Commit style: phase0:, bench:, gates:, agent:, playbook: prefixes
• When in doubt between shipping a feature and hardening measurement, harden measurement. The product is trust.