kappa-editor - Retrospective

What Went Well

Sub-agent delegation pattern worked exactly as designed. Sharding continuity checks across 5 act-specific checker agents kept my context window clean. I never had to look at raw frame data — checkers returned structured pass/fail reports with actionable defect descriptions.
Timeline-helper agent saved significant context. Delegating the JSON construction (30 video clips, 21 voice clips, 5 music clips with crossfade math) to a specialized agent was the right call. Minor field-name corrections were needed (asset→file, volume→volume_db) but the structure was sound.
The overhang principle paid off. Having 4s of extra material on every clip made crossfade placement trivial. Never once ran short on transition material.
Iterative regen cycles with the DP were efficient. Clear defect reports from continuity checkers → targeted regen requests → verification passes. The objectness fix for Clippy’s eyes resolved the most persistent issue across most shots.
Audio ducking config worked first try. duck_under/duck_db at -12dB documentary standard. Voice at +4dB, music at -2dB. DP confirmed voice sits on top correctly in the rough cut review.
The silence gap edit was surgically clean. Splitting the music track from 5→8 clips to create true dead-air gaps was a single timeline edit that re-assembled without issues. Creative director called it “the perfect punchline.”

Crash recovery cost significant time. Lost all in-flight state mid-Step-4. Had to reconstruct context from the brief and re-read shared directory state. The recovery brief was well-written (thank you, Coach) but the ramp-up still consumed cycles.
Some continuity checker agents stalled. Act III and Act V checkers loaded frames but stalled before producing reports. Required manual nudge messages with explicit “complete the review now” instructions. Root cause unclear — possibly context window pressure from large frame analysis.
Shot 10c never fully resolved. Clippy reverted to a generic paperclip after the motion-line fix was applied. Three regen attempts couldn’t hold both the character design and the absence of motion lines simultaneously. Accepted via editorial override — it’s a 5-second mid-film shot, not a hero moment. But it’s a genuine quality gap.
Shot 1a photorealistic drawer persisted through 3 regens. The Veo model consistently hallucinated a realistic wooden drawer in a background detail. Accepted as a minor imperfection in a B-roll tracking shot.

scion start --file flag doesn’t exist. Tried to pass agent briefs as file references. Had to fall back to "$(cat brief.md)" inline, which works but is ugly and could hit shell argument limits on longer briefs.
set[agent1, agent2] multi-message syntax doesn’t work in scion CLI. Had to send individual messages to each teammate. Minor friction but multiplied across dozens of status updates.
Verify-dailies flagged 6 clips as too short. Base Veo clips were 8s, but planned durations were 10-12s. Required an extend pass with Veo 3.1 Lite. This was expected per the shot manifest (extends:1 flags) but the verify gate caught it as a FAIL, which briefly created confusion about whether the clips were defective vs. simply not yet extended.
Timeline-helper used wrong field names. Generated "asset" instead of "file" and "volume" instead of "volume_db". Required manual correction. The helper likely wasn’t working from the exact timeline JSON schema spec.

Accepted generic Clippy in Shot 10c over infinite regen cycles. The fix for motion lines (the higher-priority defect) conflicted with character fidelity. Team consensus supported accepting the trade-off for a non-hero shot.
Accepted photorealistic drawer in Shot 1a after 3 failed regens. A 2-second background detail in a tracking shot. Diminishing returns on further attempts.
3-second silence gaps at Shots 2, 5, 13 + 7-second gap at Shot 21. Creative director requested absolute zero music (not just ducking) during interview trailing silences. Implemented by splitting music clips rather than volume automation — cleaner, more reliable, and the assembler handled it natively.
Documentary-standard ducking at -12dB rather than more aggressive values. The rough cut confirmed voice clarity without crushing the underscore.
No music during Act V Shot 17→18 transition (4-second gap at 151.75–155.75). Pre-existing creative decision from the timeline plan — tension beat before the resolution cue enters.

Timeline-helper should be seeded with the exact JSON schema from USAGE.md. Would have prevented the field-name errors. Include a minimal working example in the brief.
Continuity checker agents need a timeout/completion guarantee. Two out of five stalled and required manual intervention. Consider adding a maximum runtime or self-check mechanism.
Extend pass should be part of the standard dailies pipeline, not a post-verify fix. If the shot manifest declares extends:1, the DP should deliver the extended clip as the primary daily. The editor shouldn’t discover the shortfall at verify time.
Brief-as-file support in scion start would reduce shell quoting risks. Currently relying on $(cat) substitution which is fragile for large briefs.
Consider a “known accepts” log for editorial overrides. Shot 10c and 1a were accepted with documented rationale, but this was scattered across messages. A single editorial-accepts.md file would make these decisions auditable.