Milestones in Codebase Changes: A Week of Hardening CLAJAMA
We've been keeping a development log — every milestone in the codebase, why it
was built, and the native check that proves it works. This post is the public
version. Seven milestones in two days, all verified live, all built on the
principle that the cheapest correct fix is the one that never needs the LLM.
M-01 — The build modal, rebuilt the right way
The progress overlay was rendered before — a broken
document where the full-screen backdrop sat above everything, old bundled CSS
won the cascade, and mobile was unusable. There were two id="build-modal"
elements, the poller started before its own controller existed, and a stuck
build showed "Queued" with no motion.
Fixed: the modal now renders once, inside the body, after the stylesheet.
The poller defers until the page is parsed. There's a live elapsed ticker,
auto-discovery of running builds, a "needs review" state, and a proper mobile
layout. Verified live: doctype first, one modal, after the CSS link.
M-02 — "Start New Work" unblocked
Three hidden killers: builds dropped when the queue was full (no build_id),
the interview→build transition ran a blocking synchronous build (15–60
seconds in the HTTP request), and an orphaned tenant row made every project
create fail with a foreign-key error.
Fixed: the final interview answer now dispatches asynchronously and returns
a build_id instantly; full builds park in a pending queue and auto-start when
a slot frees; project creation auto-heals a missing tenant. Verified on the
live flow: thin prompt → interview → answer → build starts immediately.
M-03 — FixMap: snip, don't integrate
We evaluated FixMap — a brilliant open-source tool (MIT) that ranks which
files matter for a task before you edit. Instead of adding a Node runtime to a
Python stack, we snipped the ideas into a native tool: task_map.py gives
deterministic, no-LLM, ranked context for any task description. Credits live in
our registry — including the community member who surfaced it at 12:32 AM.
M-04 — ForkGuard engraved into the pipeline
The FixMap discipline became a runtime guard: core/fork_diagnostics.py
runs at every decision fork — build dispatch, intake→build, EOS advance — and
scans for collisions, context loss, and procedure errors. It's wired into the
30-minute integrity audit and one-command from the CLI. We refined it until
every finding was real: Jinja branch collapse + include-graph analysis killed
all ten false positives. On prod, the audit went from 32 warnings to 22.
M-05 — The Flaw Registry
Every flaw is now tracked with the check that catches it — F-01 through
F-14 — so a re-scan either proves it's gone or re-flags it. That's the loop:
scan → fix natively → re-verify with the same check → count the drop.
M-06 — Non-LLM deterministic repair
The biggest token saver yet. The top build flaws — empty tags, missing
meta descriptions, duplicate forms, machine-y titles — used to trigger a **full
LLM rebuild** (blueprint + all content + render). Now a deterministic repair
pass fixes them in milliseconds, re-validates with the same validator, and
persists the corrected pages. Verified on the live pipeline: **6 validator
errors → repair → 0 errors**, with the estimated tokens saved logged per build.
M-07 — Credits and open source
All of this traces back to open source: FixMap's ideas, the community member
who stayed up late to share them, and the tools we stand on. Our credits file
and the previous post — Big Breakthroughs Always Come From Open Source — are
the standing acknowledgment.
The point
Milestones matter less than the process that produces them: every change
ships with a native check, gets verified on the live system, and is logged. If
you're building software, keep a development log. Future-you will need it —
and the checks will keep the flaws weaned out.