** This file was created with an agent. **

#Post-Mortem: VideoproofContextual — the Model-Led Rewrite vs. the Human Finish

Written 2026-07-16. Retrospective on the VideoproofContextualActor effort, Feb–Jun 2026. Grounded in git log over the contextual paths; numbers below are reproducible from history.

#Why this document exists

The docs/planning/videoproof-contextual/ directory contains two upbeat planning artifacts from the model-led phase:

• rewrite-plan.md — a phase-by-phase plan that marks phases 1–9c “✅ Done”.
• ui-usability-plan-2026-03-16.md — a tidy usability plan for the nested template editor.

Read on their own, those documents tell a story of a smooth, phased, mostly-completed agent rewrite. That is not what happened. The honest story is: an agent produced a large scaffold quickly, that scaffold then required a long, sustained human effort to become correct and usable, and the majority of the shipped design decisions came after the model was set aside. This post-mortem records that so the planning docs are not mistaken for the whole picture.

#The numbers

Restricting to the contextual actor + template UI paths, from the first agent commit (2026-02-27) onward:

• Initial agent commit (44c6e4a7): one file, 366 lines, “new actor created by agent.”
• Agent era: 2026-02-27 → 2026-03-26 (~1 month, front-loaded).
• Human era: 2026-03-11 → 2026-06-17 and beyond (~3+ months, overlapping then fully taking over).
• Current shipped size: ~2,145 lines across actors/videoproof-contextual/ (layout.mjs 320, models.mjs 345, template.mjs 369, index.typeroof.jsx 638) plus ui-contextual-template.typeroof.jsx (473).

The human removed roughly as many lines as the agent originally added (−3,373) while adding nearly twice as much (+5,966). In other words, the final actor is not “the agent’s scaffold, polished” — it is largely a different body of code that grew on top of, and by displacing, the scaffold.

#Timeline in three acts

#Act 1 — Model-led scaffold (Feb 27 – ~Mar 12)

The agent generated the pattern compiler, selector/template engine, model definitions, and wired the actor, in labelled “Phase 1–9c” commits. This was genuinely fast at producing structure. rewrite-plan.md was written and maintained by the agent during this act, which is why it reads as a clean ledger of completed phases.

The trouble is already visible inside this act, in the agent’s own commit subjects:

• fix: correct critical bugs in template engine integration (Mar 10)
• refactor: remove ad-hoc template spec format, replace with TODOs (Mar 10)
• Fix compilePattern return value destructuring … (Mar 12)
• [videoproof-contextual] TODO: full cell rendering causes an unresponsive browser (Mar 12, human commit — a correctness/perf problem in the generated renderer).

#Act 2 — Handover / overlap (~Mar 11 – Mar 26)

Human commits begin interleaving with agent commits (“some cleaning up and refactoring”, “some cleanup”, “flesh out the CoherenceFunction”). The last agent-authored commits are algorithmic fixes to its own earlier output:

• computeFontSizeAndLayout: binary search replacing broken iteration (Mar 24)
• wordIndexToLine: binary search over sorted lineStarts (Mar 24)

i.e. the closing agent work was repairing broken iteration it had written. After Add UIArgIndexInput (Mar 26) there are no further agent commits.

#Act 3 — Human redesign & finish (Mar 17 – Jun 17)

This is where the bulk of the real design happened, and where the tone of the commit log changes from “Phase N done” to targeted correctness and modelling work. Representative human commits:

• fix cases of halluzinated propertyRoot declarations (Apr 8) — cleaning up model-invented paths that did not correspond to anything.
• AND/OR Combinators are now distinct types, not one type with a distinctive enum (Mar 29) — a core modelling decision the agent had gotten wrong.
• don't use multiple charGroups in the simple selector: the AND combinator can achieve the same thing (Mar 29) — simplifying the agent’s design.
• FIX: UIContextualTemplateContainer: inject contentsZoneElement as a widget so label, widgets order is correct (May 1).
• put template metamodel instance directly into processed properties (May 18) — a substantial rework of how the template participates in the properties system.
• Extraction of applyPresets/presets.mjs, argument-marker redesign (\1, \2), caching of compileTemplate, scrolling behaviour, colour and cell-box features (May–Jun).

The “REVIEW:” commits (Mar 17) and repeated prettier/eslint passes also show the generated code did not match project conventions (jshint headers, an eslint no-unused-vars pattern that masked unused vars containing h, etc.).

#What the model was good at, and what it cost

Good at: producing a plausible, large, phased scaffold fast. One month to ~3.7k lines of structure that compiled and ran. It correctly leaned on existing reference patterns (AxesMath dynamic types, StylePatch dispatch) that were pointed at it in rewrite-plan.md.

Cost:

• Hallucinated structure — propertyRoot declarations and paths that didn’t exist, needing explicit human cleanup (Apr 8).
• Wrong core modelling — combinators as an enum instead of distinct types; redundant charGroups in the simple selector; both later reversed by hand.
• Broken algorithms shipped as “done” — iteration bugs in layout/line computation, only fixed via binary-search rewrites at the very end of the agent era.
• Convention drift — jshint/eslint/prettier noise, awkward container wiring.
• Misleading self-reported status — the agent authored the “✅ Done” ledger, so the plan overstated completeness. Phase 10 (“Built-in templates”) is still 🔲 open to this day, and much of what was marked done was subsequently reworked.

The net effect: the scaffold was a starting point that had to be paid down, not a finished feature. The human effort (3+ months, ~6k added / ~3.4k removed) was not “polish” — it was redesign, correctness, and usability work that constitutes the actual shipped actor.

#About the two planning documents

• rewrite-plan.md is a useful historical record of the model-led phase and its intended architecture, but its “✅ Done” ledger reflects the agent’s optimistic self-report, not the final state. It should be read together with this post-mortem.
• ui-usability-plan-2026-03-16.md describes summary-first collapsed views and AxesMath-style expand-on-edit for the nested editor. Those ideas were eventually realised in the shipped code (_getSelectorSummary, UICharsSelectorSummary, getCharGroupSummaryFromModel, Collapsible usage in ui-contextual-template.typeroof.jsx) — but as part of the human-led Act 3, not as a discrete tracked deliverable. It is kept as a historical artifact of intent, not as an authoritative record of how the work was done.

Neither document should be relabelled a clean “done.” They are preserved as artifacts of the model-led phase; this post-mortem is the corrective narrative.

#Lessons (for the next model-led attempt)

1. Do not let the agent be the author of its own status ledger. The “Phase N ✅ Done” file was written by the same process doing the work; it measured activity, not correctness.
2. Treat a large generated scaffold as debt, not equity. Budget for the possibility that most of it will be removed. Here, ~90% of the agent’s added lines were eventually removed or displaced.
3. Core modelling decisions are the expensive part and the part the model got wrong (enum-vs-types, selector cardinality, property paths). Those deserve human design before generation, not correction after.
4. “Runs” ≠ “correct.” Broken iteration and an unresponsive full-render both shipped in code marked done.
5. Small, verifiable steps beat phased mega-scaffolds. The durable parts of the final actor came from many targeted human commits, each locally reviewable — the opposite of the front-loaded agent burst.

Reproduce the figures:

git log --author="Agent of Lasse Fister" --since=2026-02-27 --pretty=tformat: \
  --numstat -- '*videoproof-contextual*' '*ui-contextual-template*' \
  | awk 'NF==3{a+=$1;d+=$2} END{printf "+%d -%d\n",a,d}'
git log --author="Lasse Fister" --since=2026-02-27 --pretty=tformat: \
  --numstat -- '*videoproof-contextual*' '*ui-contextual-template*' \
  | awk 'NF==3{a+=$1;d+=$2} END{printf "+%d -%d\n",a,d}'