The janitor's source-code sibling: surveys existing code (not a diff) for smells,
dead code, duplication, and over-complexity, prioritizes by churn × complexity, and
recommends a disposition (refactor/delete/defer/accept) per finding designed to feed
/triage. Test-net status and risk-to-change are first-class so a refactor is only
recommended when behavior preservation can be proven. Read-only; the risky auto-apply
half is deliberately deferred and gated. ROADMAP item 11.
The .dc.html prototypes are parameterized components whose data-props defaults encode
the decisions the user landed on; the bundled screenshots are option-history (rejected
options + stale iterations). Read the .dc.html at its defaults as the source of truth
and confirm with the user — don't anchor the contract on screenshots. (Learned on the
ten31 CRM: a screenshot-anchored read locked a plan onto a rejected card the .dc.html
defaults had already superseded.)
Two process learnings: render hue/shade reconcile candidates as pixels
(Chrome headless screenshot of an in-context vignette) rather than relying
on hex/monospace previews; and how to absorb one owner-driven accent in a
document-as-is extract (single canonical hue, distinguish overlap by
treatment not a second color, surface the collision before picking).
Three process learnings from the 2026-06-19 Ten31 mobile round-trip:
- Output read via the DesignSync MCP is a multi-file app on Claude Design's
own template runtime using CSS custom properties (incl. a dark/light
switch) — qualifies the seed note about "inline hardcoded CSS" (that's the
export-bundle path, not the MCP project-read path). Read the shell first,
then each sub-app's --var block + DCLogic for derived-field formulas.
- DesignSync reads into context only (no bulk download / no screenshots):
byte-capture the representative source + a manifest README; use
`jq -r .content` on the harness-persisted result for large files.
- A redesign can return more scope than the brief asked for (a light theme);
record it but write scope-expanding additions as proposed-not-canonical
pending an explicit owner decision, surfaced as an A/B reconcile question.
Keep the investigation and the judge's decision rigorous and
fact-based, but render everything shown to the owner — both debate
sides and the rationale — in plain language. ESCALATE now surfaces an
explicit For it / Against it / Judge's lean pair.
Drop the finished build narratives (durable record lives in ROADMAP
items 8/9/10), fix the stale "built this session" on onboarding-tester,
and surface the /adjudicate first-run drop-bias calibration as the top
next step.
Parked P2/P3 technical items accumulate faster than I can judge their
necessity. /adjudicate runs a grounded per-item debate (investigator →
build/drop advocates → judge) over a repo's ROADMAP and routes each to
DROP / DO / ESCALATE, so I ratify decisions instead of researching them.
Recommend-only in v1; verdict autonomy is gated by blast radius, not
priority. ROADMAP-only input — nudges /triage rather than reading the
raw inbox.
Pair the existing /compact keep-alive line with a paste-able opener for a
fresh context window. Constrained to a pointer into AGENTS.md Current state,
not a state payload, so it stays safe to lose.
First extract->reconcile run (recap) surfaced generalizable process
learnings: harvest the inventory with grep frequency tables in the main
thread (counts are the reconcile evidence), disambiguate near-duplicates
with frequency plus an external anchor, present conflicts as
recommended-first A/B/C forks with value previews, and treat the code
itself as the inspiration for a document-as-is extract (skip BRIEF and
_imports, write provenance). Also refine the Extract phase to enumerate
every styling surface and read the brand mark before the CSS.
Drop two recap items (daily-digest SMTP shipped in 0.2.158; mobile
scroll-to-top already tracked in recap Current state). Retag the
"run janitor on all projects" item from (?) to (standards).
Revise ROADMAP item 9 and the INBOX Path-2 item: split the harness into
Stage 1 (Path 1, no payments, unblocked) and Stage 2 (Path 2, full
buyer-pays on regtest), gated on keysat's greenlit payment_providers:write
scope + network gate + sandbox flag. Drops the earlier 'operator
pre-connects' framing now that scoped agent connect on regtest is the plan.
Global fleet agent that walks a product's published docs as a literal
newcomer, never reading its source, reports every doc gap, and emits a
publishable walkthrough only on a fully clean run. First instantiation:
keysat SDK integration. Register in fleet list, log ROADMAP item 9 and a
Current-state note, resolve the originating inbox item, and capture two
follow-ups (Path 2 payment demo, operator-onboarding agent).
Keep "Claude Design" only as the named default visual tool (like naming
Figma); generalize the spots that assumed the coding harness is Claude:
- Phase B framed as "the cloud tool" (Claude Design as default; adapt for Figma)
- export named the "coding-agent handoff bundle" (Claude Design's term noted)
- token extraction points at Style Dictionary / any harness skill, not Claude Code
- BRIEF prompt block paste target generalized
Keysat