The 1.63 arc closes. Chapters one and two landed at 1.62.1; chapters
three and four landed on main just before this; the minor is the story
finishing, and its own contribution is the closing condition the
plan committed to: the gate goes live.
Added#
trazum.bench.json, committed, and CI held to it. The Build-and-tests job now runs the ratio gate against the recorded baseline with a stated factor of 3× — generous on purpose: a tripwire for the quadratic-regex class of regression, not a micro-benchmark. Recorded and immediately cross-validated on a machine whose CPU changed between the recording run and the gating run: wall clocks moved ~1.5×, ratios held (5.47→5.33, 11.08→11.14) — the machine cancels out, measured rather than claimed. The plan's named risk stands: a workload whose ratio proves too noisy on shared runners loses its gate loudly in the release notes, never left flaking.Memory holds a line — chapter four of the 1.63 arc. The 25MB usage log profiles within a stated 384MB heap ceiling, asserted in the suite the only way that is honest on any machine: the probe runs in a child process with V8's old space capped at the line, so "fits" is enforced by the engine rather than inferred from an RSS reading a memory-rich runner would inflate by collecting lazily. Deterministic 25MB log (the fuzzer's LCG, a fixed clock), every line parsed, none skipped. The line was proved to bind in both directions before it was trusted: the same probe dies at 64MB, and a deliberate memory hog dies under a 48MB cap inside the suite itself, so the flag is not decoration. Observed peak RSS for the run: ~158MB, which is the headroom the release notes publish next to the promise. Moving the line is a release-notes decision, the same as the token band.
The refusal ceiling — chapter three of the 1.63 arc. Above 400,000 characters an input that claims to be a prompt is refused with the size and the limit named, never ground through. The number now lives exactly once, as
MAX_INPUT_CHARSin@trazum/core; the two web routes, the share endpoint and the MCP server — which each carried their own400_000, agreeing by coincidence — now derive from it, and a new suite guard holds every other prompt-ceiling constant in the repository to deriving rather than restating (with three named non-door exceptions: the localStorage history bound, the prompt library's row bound, and a name length — each with its reason written beside it, because "it is different" is the sentence that stops being true quietly). The CLI's prompt doors —optimize,check,diff,eval,prune,semantic, the prompt half ofroute, and every directory walk that feeds prompts to the optimiser — hold the same line; a global--max-input <chars>raises it deliberately, and lowering it is equally legitimate. Logs and documents are never held to it:profileandconformread inputs fifty times this size by design, and the null at those call sites is written out so a reader sees the decision.
Changed#
- One CodeQL query excluded, out loud and owner-approved:
js/file-access-to-http. Merging the refusal ceiling surfaced four medium alerts — the three provider calls inllm.tsand the counter intokenizer.ts, each "file data in an outbound network request". That shape is exfiltration in most software; here it is the contract:--exact-tokensand--llmsend the prompt to the API the caller configured, opt-in by flag, gated on the caller's own credential, documented as doing exactly that. The protection the query wants exists in the security suite as tests — which modules may reach the network at all, and that no command calls out unasked. The exclusion was put to the repository owner as a question and approved before it landed; it lives in.github/codeql/codeql-config.ymlwith the full reasoning attached, reviewed and diffable rather than a dismissal buried in a settings page. Every other security-extended query stays on.
Release mechanics#
- Five manifests,
@trazum/corein three places, lockfile at 1.63.0; README Action pins advanced to214aa43(1.62.1).