yoloyolo/.planning/RETROSPECTIVE.md

Retrospective

Milestone: v1.0 — MVP

Shipped: 2026-03-26 Phases: 4 | Plans: 11 | Timeline: 3 days (2026-03-24 -> 2026-03-26) LOC: 3,254 Go | Files: 84 modified

What Was Built

• Config-driven packet classifier with 12 protocol rules and 4 hash-bucketed unknown classes
• Live packet capture via pure Go go-pcap with privilege detection and non-blocking channel pipeline
• Additive synthesis engine: per-class oscillators, EMA amplitude smoothing, constant-power stereo mixing
• MP3 encoding via embedded LAME, zero-packet guard
• End-to-end pipeline with clean Ctrl+C shutdown
• BPF filter and offline pcap file sonification with timestamp-based windowing

What Worked

• Risk-ordered phases: hardest foundation first (capture), then synthesis in isolation, then integration — no phase blocked on another's bugs
• TDD approach in executor agents caught signature mismatches early
• Pure Go capture layer (go-pcap) avoided libpcap dependency headaches
• Hand-rolled synthesis kept dependencies minimal and code understandable
• Wave-based parallel execution for independent plans

What Was Inefficient

• Some SUMMARY.md files had empty one-liner fields, causing noisy milestone extraction
• Phase 4 research could have been lighter — BPF/pcap APIs were straightforward

Patterns Established

• Ordered []Rule slice for classifier (first-match-wins, extensible)
• io.Writer injection for testable stderr output
• Buffered channels between pipeline stages (512-1024) to absorb bursts
• Package-level globals for Cobra flag binding
• Timestamp-based windowing for non-live sources

Key Lessons

• Embedding C source (go-lame) is a good tradeoff: CGo at build time only, no runtime dependency
• go-audio/wav was unnecessary — writing PCM bytes directly to LameWriter is simpler
• Hash-bucketed unknowns (4 classes) are sufficient for audio distinction without k-means complexity

Milestone: v1.1 — Custom Sound Mappings

Shipped: 2026-03-26 Phases: 3 | Plans: 6 | Timeline: 1 day (2026-03-26) LOC: ~4,675 Go (+1,421 from v1.0) | Packages: 7

What Was Built

• Four waveform types (sine, square, sawtooth, triangle) with bandlimited additive synthesis
• TOML config system: auto-discovery, partial overrides, unknown-key validation, --config flag
• User-defined [[rules]] classification rules with first-match-wins prepend semantics
• Auto-frequency assignment (FNV-32a hash) for custom class names
• --print-config flag with commented TOML output and source annotations
• LoadResult struct pattern for clean config-to-pipeline data flow

What Worked

• Incremental config extension: Phase 6 built the config package, Phase 7 extended it cleanly
• TDD plans (type: tdd in frontmatter) produced higher-quality code with fewer regressions
• FNV-32a frequency assignment was verified experimentally during research before planning
• Worktree isolation for parallel executor agents prevented merge conflicts
• Reusing existing patterns (Rule struct, NewClassifier injection, Cobra flag-on-root) kept code consistent

What Was Inefficient

• SUMMARY.md one-liner extraction continued to be noisy — summary-extract needs improvement
• Phase 5 could potentially have been merged with Phase 6 (waveform + config together)

Patterns Established

• LoadResult struct for multi-value config returns (extensible without breaking callers)
• Pointer fields (*float64, *string) for partial TOML override semantics
• FNV-32a hash for deterministic resource assignment from string keys
• Flag-on-root pattern for early-exit operations (--list-interfaces, --print-config)
• addAutoFreqEntries() pattern: fill gaps in config before merge

Key Lessons

• BurntSushi/toml Undecoded() works with array-of-tables (verified experimentally)
• Manual string building beats TOML encoder when you need inline comments/annotations
• Config extension is smooth when the original Load() was designed with clean boundaries

Cross-Milestone Trends

Metric	v1.0	v1.1
Phases	4	3
Plans	11	6
Days	3	1
LOC	3,254	4,675
Avg plan duration	~8 min	~5 min