A Narration Script for Every Feature Screencast
One fresh agent per episode, each writing an anchored script the replay pipeline can film
Gantry's screencasts are narrated tours of what a real run did — a documentary voice over a recording of the build, with a small framed panel that fades in to show the actual code, plan, or test command a beat is about. Each episode needs a written script: the narration, plus invisible anchor comments that pin sentences to moments in the run and overlay directives that name the artifact to show. Writing those by hand, one per feature, is slow and drifts out of a shared register.
This job turned that into a fan-out. An episode manifest lists every screencast; the run starts one fresh agent per episode, and each writes a complete, concise script for its own episode against a validator that checks the anchor and overlay grammar. What the run produced is a corpus of ready-to-film scripts covering Gantry's features, from the plan cards to harness selection to merge reconciliation.
Build
The map cut the work into one independently gated script per episode, so a single script the validator kept rejecting could be set aside without holding back the rest.
The decomposition here is the map itself: each episode in the manifest is its own unit, carried by a fresh agent that writes one script and is judged by one validator run over that script alone. Two properties fall out of cutting it that way, and both showed in this run. Because a unit is skipped when its script already exists and passes the validator, most episodes never re-ran — the corpus was largely in place, and the run confirmed it rather than rewriting it. And because each unit is gated on its own, one failing episode does not block the others.
That second property is what the run leaned on. The episode for the bare `gantry` chat command is where the difficulty lived: its agent wrote a script, the validator rejected it, a fix ran, and the validator rejected it again — across repeated passes the execute-then-fix cycle never turned the gate green. The recorded reason is precise: the script opens with an anchor naming a run stage the validator does not recognize, so every version tripped the same check. The run closed accepting that one episode as unfinished, its ledger rows removed so the count reflects the scripts that actually passed, and left the broken script in the tree for a person to fix later.
The two episodes that did run fresh — the pluggable-harnesses script and the usage-limit-reset script — went through execute and gate once each and passed, needing no repair. Nothing forced a re-plan; the boundaries drawn by the manifest held, and the gate did exactly the isolating job the per-unit cut exists to provide.
Feature
Before this, screencast narration was hand-written per episode; the run leaves a validated corpus of anchored scripts in the tree, with the one episode the validator refused still standing broken for repair.
A screencast script is not free prose. The replay pipeline reads anchor comments to synchronize narration to the recorded run and overlay comments to stage the framed panel, and a dangling anchor or an unknown stage label fails the validator the same way a code test fails. Before a script exists and passes that check, an episode cannot be filmed. The purpose of this run was to bring the episode manifest and the script corpus into agreement: a valid, anchored, concise script sitting next to every episode the manifest names.
Standing in the tree today, that corpus is present under the screencast scripts directory, one Markdown script per episode, each with its front matter, its anchor track, and its overlay directives. Running the validator against a completed one — the pluggable-harnesses script — still reports it clean, with its full set of anchors, so the grammar these scripts were written to has not drifted out from under them.
The exception is exactly the one the run flagged. The bare-`gantry`-chat script is still in the tree and still opens with the anchor naming an unrecognized stage — the same defect that kept it from passing during the run. It remains the one episode whose script is not yet filmable, left in place deliberately rather than dropped, so the gap is visible to whoever fixes it rather than hidden.