adhoc (Preston's advisor)

What Went Well

Parallel investigation with tool-maker was highly effective. I analyzed the assembly source code and theta’s clip properties while tool-maker prepared a briefing on the 5 original root causes. Converging findings independently gave high confidence in the diagnosis.
Direct clip inspection sealed the diagnosis. Using ffprobe to check keyframe positions in Veo clips (0.033s, 8.033s) and then running the exact trim command to observe -1.0s packet timestamps made the root cause undeniable.
Cross-agent information gathering worked well. Messaging theta-techlead with specific diagnostic questions got comprehensive answers quickly, including the critical detail that 5 clips used the old concat approach.
Tool-maker implemented and deployed the fix quickly once we agreed on the root cause. The trimAndStandardize() function is clean — single re-encode pass, no extra overhead.

ffprobe wasn’t on PATH. I wasted several attempts checking stream properties before discovering /workspace/tools/bin/ffprobe existed. Should have checked /workspace/tools/bin/ immediately.
Context compaction mid-investigation. The conversation was compacted during the sync bug investigation, requiring reconstruction of analysis state from the summary. Some nuance was lost.
Initial theory about -c copy was partially wrong. I initially thought the A/V offset direction would be “video leads audio” based on keyframe-before-seek behavior, but the actual FFmpeg behavior with negative timestamps and MP4 edit lists is more nuanced. The empirical test (running the actual trim command) was what confirmed the direction.

Scion CLI quoting issues. set[agent] syntax fails in zsh due to bracket globbing. Had to use single quotes around the recipient. This happened twice before I remembered the workaround.
No ffprobe in PATH. The genmedia tools bundle their own ffmpeg/ffprobe in /workspace/tools/bin/, but this isn’t documented in the agent templates or CLAUDE.md. Every agent that needs to inspect media will hit this.

Asked tool-maker to make the edit, not me. Preston’s instruction was “compare your findings before agreeing on any edit (and who makes it).” Since tool-maker owns the genmedia codebase and had full context on all 5 prior root causes, they were the right person to implement the fix. I focused on verification.
Combined trim + standardize in one pass rather than adding a separate re-encode trim step. This avoids double re-encoding and is simpler. The alternative (re-encode trim → standardize) would have been correct but wasteful.
Did not ask theta to re-do the 5 buggy extends immediately. The source_in/source_out trimming in the timeline hides the duplicated footage, so the priority was fixing the sync bug and re-assembling. The extends can be re-done in a follow-up if quality is impacted.

Add /workspace/tools/bin to PATH in agent templates. Every agent that touches media will need ffprobe. Currently they have to discover it by listing the directory.
The extend workflow needs a harder gate. Two bugs (concat double-count and now A/V sync) trace back to the extend workflow. Consider: (1) verify-dailies should flag clips with video start_time != 0.000 as potential concat artifacts, (2) the playbook should mandate genmedia-video extend output be used directly (mv, not concat) with a checklist item.
Playbook should document the assembly subcommand explicitly. Theta used timeline, eta used render — both work but have different capabilities and fix histories. Teams should know which to use and why.
Cross-agent debugging sessions benefit from a shared scratch file. When multiple agents investigate the same bug, a shared markdown file (e.g., scratchpad/bugs/av-sync-investigation.md) with timestamped findings would prevent duplicate work and make the investigation auditable.