05 / Maintenance
Who edits it matters.
So does what remains.
Simpler build ≠ less shipped JS
Agent has no normal Astro build/runtime requirement or React install, and no Astro hydration islands. DocSearch is a documented Preact compatibility exception for modal, keyboard, focus, ranking and history state. Mermaid, SDK, auth and crypto libraries remain. Production inline controllers and external integrations are explicitly retained, not claimed newly invented vanilla code.
Total rebuilt Agent JS is 6,082,284 bytes, compared with original 6,073,203 bytes: slightly larger. There is no demonstrated overall client-JS reduction. These are uncompressed asset totals, not per-page downloads. All versions retain production CSS/images/fonts.
Inspectable, but substantial
Agent maintains 1,347 wrappers, 1,881 shared HTML includes, 5,990 converted fragments and six templates/helpers. Thirty-seven readable JS files supply all 21 entry controllers. Its locked dependency set has 187 packages and 228,622,975 bytes of installed regular files.
Faithful’s root lock has 188 packages, plus the separate MDX renderer and other inputs; upstream’s Bun monorepo scope differs. Equivalent total install/download footprints were not measured, so no percentage saving is claimed. Agent keeps 19 original CSS files and removes only audited, unreferenced JS from served output.
What would we maintain?
If we were the existing Capgo organisation today
Staying with upstream avoids migration and new certification work, and retains its deployed translation/service system. That is an organisational-risk argument, not a demonstrated development-loop advantage. Team familiarity is unknown, private workflows remain uncertified, and comparable Astro dev/HMR/edit measurements are missing.
If we were committed to Nift, with humans + agents
We would prefer faithful when retaining MDX/content authoring matters. Its captured corpus, source-backed metadata and shared composition are demonstrated; ordinary real-edit family medians are 4.28–6.41 s, with explicit targets at 2.07–3.71 s here. Schema/discovery and new style/asset-generation conveniences still need acceptance or further tooling.
If we were designing for agent-mediated implementation
We would prefer Agent for that premise: explicit source mappings, inspectable markup, readable controllers and parity gates, with normal real-edit family medians of 1.04–1.09 s and targets at 0.86–0.91 s. Those timings support cheap build feedback; they do not measure total validation latency, agent cost or editing reliability. Humans still direct/review, and shared changes require broader builds and checks.
A preference for a Nift engineering workflow and caution about a real production migration can coexist. We cannot decide developer experience from the Astro full build alone.
If migration weren’t part of the equation
Assume all three implementations already exist, are equally production-certified, have equivalent access to the required backends, and the team can work effectively with either stack. Migration, switching cost and incumbency are excluded. The following are engineering preferences based on this corpus, not measured long-term productivity results.
Humans + agents: faithful Nift
We would choose faithful Nift. It keeps authored MDX/Markdown, frontmatter and shared composition readable for people while giving agents explicit mappings and parity gates for corpus-wide work. Measured ordinary real-edit family medians are 4.28–6.41 s and targeted medians 2.07–3.71 s. We prefer that combination of familiar content authoring, programmable preparation and a small measured build budget over maintaining converted HTML by hand.
Astro still provides more complete collection validation/discovery, typed component and scoped-style compilation, and packaged build adapters. Faithful’s bounded adapters and explicit route registration need care; retained compiled browser assets make new browser behaviour and CSS changes less direct to debug and regenerate. Those are real maintenance costs we would accept for this workflow, not capabilities we claim to have finished. Equal backend access removes the worker deployment advantage; translation remains a separate system in either case. Dependency totals are not comparable across the installs, and retained browser libraries remain substantial. Astro edit/HMR was not measured, so this preference does not claim that Nift beats Astro’s day-to-day latency.
Agent-primary: Agent Nift
We would choose Agent Nift. Deterministic source mappings, shared includes and readable controller ownership make routine edits and debugging explicit. Agents can make mapped bulk changes while humans review intent, diffs and acceptance. Its measured ordinary real-edit medians of 1.04–1.09 s and targeted medians of 0.86–0.91 s support repeated build feedback. Independent JS bundling and the absence of a normal Astro build or hydration islands simplify the build architecture; route, dependency, parity and representative browser gates supply concrete checks.
This choice accepts a large fragment corpus and an explicit workflow for new routes, styles and shared changes. It does not imply less shipped JavaScript, fewer total dependencies, exhaustive validation, or proven lower agent cost. Browser SDKs and the DocSearch Preact exception remain. Targets alone do not cover all consumers of shared changes; broader builds and browser checks remain necessary. The documented faithful restoration issue also shows why parity checks belong in the maintenance loop. With incumbency neutralised, we prefer Agent’s inspectable HTML and controller pipeline for agent-led implementation, and faithful’s authored content model for human + agent work.
Maintenance means changing things
For shared changes, build the complete affected set and inspect relevant desktop/mobile behaviour. Intentional new content/design requires updating the golden policy, not treating old byte equality as acceptance. We did not measure source-location counts or human/agent effort for these scenarios.
Content: a concrete convenience, not a capability verdict
Upstream’s glob-loaded blog collection declares Zod fields, supported locales, dates and tag normalisation; routes query the collection. Adding a published source entry is discovered by that pipeline. Faithful retains authored MDX/frontmatter and discovers English blog files for listings, but new rendered routes also need body/metadata mappings and Nift registration. Its narrow YAML adapters do not apply the complete upstream collection schema. Agent edits mapped converted HTML and explicitly registers routes.
These are current migration/tooling gaps and different workflows. Nift already composes authored sources, MDX, data, templates and dependencies; a schema validator or route generator can run in preparation. No fundamental content-modelling limitation was demonstrated.
Composition: compare the operation
Capgo uses typed prop declarations, named slots and compiler-scoped styles. Faithful’s registered component adapter already handles parameters, children and named slots; Agent uses shared Nift includes/functions and readable browser controllers. Exact rendered output demonstrates sufficient composition for this corpus. Arbitrary new Astro syntax or freshly scoped CSS is not guaranteed by the bounded migration adapter. Upstream packages that compiler workflow; this is a concrete convenience difference, not proof that Nift cannot compose components.
Localisation: the translation worker does the work
The shared Astro configuration builds English only. Nine advertised language choices point to a separate Cloudflare edge worker, which translates the English origin, stores page/state records, rewrites navigation, language metadata and hreflang, and implements pending-translation retries/fallback. Shared copy is English with generated translation contexts. The Nift versions preserve English output and language links; neither deploys or certifies that worker.
Adding a locale requires changes to web/docs locale definitions and the worker’s supported set/configuration, then deployed validation. We did not measure comparative file counts or effort. The existing worker is an upstream implementation-coverage advantage; the claim that Astro’s i18n model is inherently better for Capgo is withdrawn.
Integrations: turnkey adapters versus programmable seams
Astro’s configured MDX/Starlight, DocSearch, sitemap, icon and Tailwind/Vite packages offer concrete ready-made conveniences. Nift does not require a dedicated framework plugin for each underlying system. Faithful’s hook runs Node content, code and metadata preparation, processes recorded API JSON and consumes generated files with dependency tracking. Agent’s hook independently bundles browser packages with esbuild; its allowlisted provider server composes compatible HTTP services. These are demonstrated orchestration seams.
Algolia, Mermaid, Supabase, metrics, pricing, embeds and browser SDKs connect through ordinary packages, configuration, HTTP or browser code. Translation and signing/device services require separate deployed backends. Retained sitemap/LLM/raw-Markdown assets are not proof of automatic freshness after new content. No inspected integration established an intrinsic inability of Nift to connect; several ports and live workflows remain incomplete. Native/FFI was not exercised and is not counted as evidence here.
Available adapters are distinct from general interoperability. Community size, documentation quality, team familiarity and long-term editing productivity were not quantified. Every identified Nift drawback is classified as an experiment coverage gap, migration tooling gap or different workflow in the source-backed maintenance investigation. It inventories integration families and traces concrete editing operations; no implementation was changed for this reporting correction.