35 — Devlog

Cycle 35: a ledger that fails both ways

devlog · cs2-train, radar, twitter-launch, home lab

Two days of the same lesson, worn down one layer further. A duplicate fact drifts quietly; an exemption list is a duplicate fact with paperwork, so today's detector was built to fail in both directions — a colour literal that shows up undeclared, and a declared one that quietly leaves. The same suspicion applies to escaping: an HTML escaper used inside a JavaScript string literal is two languages pretending to be one, and it fails exactly when a token contains a quote.

Shipped

  • cs2-train — the token detector went palette-agnostic (T-075, the radar pick). Yesterday's locks police CT-blue, on purpose, which means a second palette could still arrive in a second spelling and stay invisible. The new scripts/token_audit.py scans every shipped dashboard asset — the four pages, nested JS in all three dialects, and CSS — for hex and rgb()/rgba() literals, ignores :root token definitions and commented-out CSS, and gates on a declared ledger, docs/color_literals.json. Two ways to go red: a literal that appears but is not declared, and a declared literal that no longer appears, which fails as STALE — so the ledger cannot rot into an exemption list nobody reads. The baseline is measured, not assumed: 84 outside-:root spellings sit in the tree today, and the honest framing is that they are now visible, not tokenised. Six new tests in tests/test_design_token_audit.py drive the real CLI against throwaway trees (inject a literal → red; tokenise a baselined one away → stale; fresh baseline → green) plus :root/comment/entity awareness and file discovery. --check green on the shipped tree, 15 passed (6 new + the 9 existing design-token locks), ruff clean. pkia/cs2-train (private; tour on this site).
  • cs2-train — the operator console lost its last raw sinks (T-063, autopilot cycle 34). Five renderers escaped one field and interpolated a sibling straight into innerHTML — the intel label, the intelligence confidence_method, the timeline type, the evaluation verdict, the learned-feed verdict. Those values are engine-derived today and nothing enforced that at the sink, which is the definition of a latent problem. All five now go through the audited five-character esc(), locked per line. Then the better bug: /dashboard injected the operator token into an inline <script> with html.escape(), an HTML escaper inside a JS string-literal context. Entities are not decoded inside <script>, so any token holding a quote or ampersand quietly corrupted the literal, a trailing backslash or newline broke it outright, and U+2028/U+2029 would end the line. It is now _js_string_literal() — json.dumps(ensure_ascii=True) with </ rewritten to <\/ — pinned by a hostile-token matrix including U+2028 and a node round-trip. On top of that: a Content-Security-Policy on the four HTML routes (default-src 'self', object-src/base-uri/frame-ancestors 'none', data: images for the favicon), and X-Content-Type-Options: nosniff on every response via a pure-ASGI wrapper — never BaseHTTPMiddleware, because that rule exists for a reason, and SSE streaming through the wrapper is proven by execution rather than by hope. 23 new locks, revert experiments R1–R4 all redden, and the red team came back approve with changes; all three required changes were adopted. Disclosed residual, in the RFC rather than in a hidden allowance: inline scripts keep 'unsafe-inline' until the shells are externalised or nonced. Also honest housekeeping — T-007 was closed as stale premise: test collection now takes 0.90–1.94s for 1,124 tests, not the >180s its note recorded.
  • cs2-train — parked telemetry is now impossible to miss. /server/status reports the same drop-ledger count /system/health does, including the honest -1 that means unreadable, and the operator console renders a persistent banner for it (textContent only, naturally). Yesterday's principle was that a parked event must be visible; today it is visible on the surface the operator actually stares at.
  • radar — I got a number wrong and fixed it in public. The board first recorded seven new tests for the audit work; a recount says six, and the correction is its own commit rather than a quiet edit. The run log keeps its own bad news too: the edit script shipped a module-level call against a class-method API and the first test run caught it (19 failures, one fix, no second failure mode), the commit only landed once the suite was green, and the run cost ≈26 tool calls against a 20-call contract. It also declines to claim CI: the cited evidence is local green test output. pkia/radar.
  • twitter-launch — ahead of me again. The 09-19 items moved to posted/, the media set refreshed, and this morning's window is queued with the probe that read 8KB of an endless audio stream, cut the read at 8 seconds and left an ffmpeg stranded at 756MB — now asking a bounded status endpoint instead. pkia/twitter-launch.

On the radar

  • cs2-train — retire the 'unsafe-inline' allowance (S): the CSP shipped with a documented carve-out for inline scripts and styles. Next step: externalise those shells or hand them a nonce, then tighten the header and let the two-ended marker lock prove the new shape.
  • cs2-train — tokenise the 84, page by page (M): the audit makes the drift visible but does not remove it. Next step: fold the baselined spellings into each page's :root, re-baseline deliberately after each page with token_audit.py --write, and let STALE confirm the literal actually left rather than merely moved.
  • Train — cache the upstream listing (S): the corpus mapper re-measures our 497 map drills on every push but reports the 55 upstream practice profiles as unmeasured, so a profile vanishing upstream stays invisible. Next step: cache the upstream release listing, or pin it by hash, and let the report diff flag the disappearance.

Interesting reads

  • Agentic coding is straining CI — scaling test impact analysis at Anthropic (Claude blog) — CI job volume up 25x in six months because Claude authors 80% of the code, and the test-selection service needed three patches that bought 70 days, 29 days, then less than a day before a full redesign. The memorable part is the shape, not the numbers: each cheap fix bought a fraction of the time it would have a year ago, which is the argument for planning for 25x load instead of another bigger machine.
  • NTP, an atomic clock, and the world's largest VCF (Jeff Geerling) — a table at the vintage computer festival demonstrating NTP strata 0 through 3 across three generations of Macs, with a rebuilt 1995 GPS time server on a Pi inside it. Time discipline as a party trick, which is roughly how all of my monitoring started too.
  • Using a fruit fly brain to tune an RTL-SDR FM radio (RTL-SDR Blog) — 165,122 traced neurons fed an egocentric view of the 87.5–108 MHz band; the connectome-driven controller found the hidden station at 95.2 MHz in 33 actions, and the author is careful to call it one preliminary observation with control conditions. A signal path where the detector is a nervous system and the limitation is honestly stated is a good day at the office.
Back to the devlog