CLAUDE.md is at 90.1% of its hard cap and the auto-loaded closure is 47% over budget; the previously-trimmed fragments have regrown

[d-morrison]

scripts/check-context-closure.py reports two conditions on main at 491906bf, and its own output says to treat the second as the moment to act rather than the moment to note it. Filing so the observation is durable, per report-mistakes-proactively.

Measured

$ python3 scripts/check-context-closure.py
/home/user/ai-config: 74 file(s), 1,173,585 bytes (~293,396 tokens at 4 B/token)

  largest 10:
    135,408  11.5%  d0  CLAUDE.md
     87,448   7.5%  d1  shared/workflow/ardi.md
     80,901   6.9%  d1  shared/principles/fail-fast.md
     65,527   5.6%  d1  shared/workflow/address-every-comment.md
     55,158   4.7%  d1  shared/workflow/algorithmatize-checks.md
     49,984   4.3%  d1  shared/workflow/fully-clean.md
     48,639   4.1%  d1  shared/workflow/metacognitive-monitoring.md
     41,202   3.5%  d1  shared/writing/fact-check-prose.md
     33,132   2.8%  d1  shared/workflow/sync-with-main.md
     28,693   2.4%  d1  shared/workflow/pr-on-claim.md

  OVER BUDGET by 973,585 bytes (~243,396 tokens); budget is 200,000.

  CLAUDE.md: 135,158 characters against the harness's 150,000-char cap (90.1%), 14,842 to spare.
  WARNING: within 10% of the cap. This file has grown by thousands of characters in a day
  before, so treat this as the moment to trim rather than the moment to note it.

Measured 2026-08-10. Both figures are volatile by construction, so re-run rather than quoting these.

Why this is not already covered

Three neighbouring issues, none of which covers this condition:

• Trim the CLAUDE.md closure: move case records out of ai-config-only shared fragments #1193 (closed) trimmed case records out of the ai-config-only shared/ fragments. That worked, and the six largest d1 fragments above have since regrown past where the trim left them -- ardi.md, fail-fast.md, and address-every-comment.md all already have .cases.md companions and are still 66-87 KB each. So the trim is a recurring obligation rather than a one-time fix, and nothing currently tracks the recurrence.
• Trim case records from lab-manual-shared fragments (needs lab-manual coordination) #1191 (open) covers trimming the lab-manual-shared fragments, and is blocked on external coordination. It is one source of relief, not the condition.
• The file-size instrument exempts every auto-loaded file, so CLAUDE.md (1,072 lines) is unmeasured while memories/ is gated #897 (open) covers the file-size instrument exempting auto-loaded files. That is a gap in measurement; this is the measurement having been taken and read.

The sharper half

The two numbers fail differently and only one of them is graceful.

Being 47% over the token budget degrades quality: the closure still loads, and rules compete for attention. Crossing the 150,000-character cap is a hard harness limit, and CLAUDE.md has 14,842 characters of headroom -- which this repo has consumed in a single day before, per the instrument's own warning text.

So the priority is CLAUDE.md specifically (d0, 135,408 bytes, 11.5% of the closure), not the corpus total.

Suggested direction, not a decision

Deliberately not proposing the trim itself, since which sections earn their place is a judgment call for the maintainer rather than something to settle in a filing:

1. Apply Trim the CLAUDE.md closure: move case records out of ai-config-only shared fragments #1193's pattern to CLAUDE.md itself -- it already has a CLAUDE.cases.md, so the mechanism exists and only the sweep is missing.
2. Consider whether the recurrence wants an instrument rather than a periodic issue: a CI step that fails when d0 crosses some threshold below the cap would make this self-reporting, per algorithmatize-checks. check-context-closure.py already computes the number; nothing gates on it.

Not a defect: the unbalanced-fence warning

The same run reports CLAUDE.md (1 unclosed fence marker(s)). Checked before filing rather than folded in: the script's own comments at scripts/check-context-closure.py:161-175 name this repo's CLAUDE.md as the known case, explaining that it "contains same-length nested fences" and that surfacing them as a warning rather than swallowing to EOF is the deliberate trade for a measuring instrument. A naive count agrees the file is balanced (16 line-start fence markers). No action.

[d-morrison]

Re-measured two days on. The CLAUDE.md cap half of this issue has been fixed; the closure-total half has gotten worse.

$ python3 scripts/check-context-closure.py
/Users/ezramorrison/Documents/GitHub/ai-config: 81 file(s), 1,215,880 bytes (~303,970 tokens at 4 B/token)

  largest 10:
     89,175   7.3%  d1  shared/principles/fail-fast.md
     87,448   7.2%  d1  shared/workflow/ardi.md
     83,095   6.8%  d0  CLAUDE.md
     68,089   5.6%  d1  shared/workflow/address-every-comment.md
     58,812   4.8%  d1  shared/workflow/algorithmatize-checks.md
     52,138   4.3%  d1  shared/workflow/fully-clean.md
     48,639   4.0%  d1  shared/workflow/metacognitive-monitoring.md
     41,202   3.4%  d1  shared/writing/fact-check-prose.md
     33,132   2.7%  d1  shared/workflow/sync-with-main.md
     32,636   2.7%  d1  shared/workflow/pr-on-claim.md

  OVER BUDGET by 1,015,880 bytes (~253,970 tokens); budget is 200,000.

  CLAUDE.md: 82,893 characters against the harness's 150,000-char cap (55.3%), 67,107 to spare.

Measured 2026-08-12. Volatile by construction, so re-run rather than quoting.

The two halves moved in opposite directions

	2026-08-10 (491906bf)	2026-08-12	delta
CLAUDE.md (d0)	135,408 B	83,095 B	-52,313
against the 150k cap	90.1%	55.3%	-34.8 pp
closure total	1,173,585 B	1,215,880 B	+42,295
files in closure	74	81	+7

So the priority this issue named --- "CLAUDE.md specifically, not the corpus total" --- was addressed, and the cap warning has cleared with 67,107 characters of headroom. CLAUDE.md has fallen from first to third largest.

The corpus total went the other way. It grew 42 KB in two days while the one file being actively trimmed shed 52 KB, which means the non-CLAUDE.md closure grew by roughly 95 KB over the same interval. Seven new files entered the closure.

What that suggests about the recurrence

This is evidence for point 2 of the issue's own suggested direction, and against point 1 being sufficient on its own. Trimming CLAUDE.md worked and did not slow the total, because the growth is distributed across d1 fragments that each look individually justified. fail-fast.md and ardi.md are now both larger than CLAUDE.md, and both already have .cases.md companions --- the same regrowth-past-the-trim pattern this issue records for #1193.

A per-file d1 threshold would be the instrument that catches this, since the total is a sum nobody owns while each fragment has an author who can be asked.

Why it matters more in-repo than the token figure suggests

Worth stating explicitly since it doubles the cost and is easy to miss: working inside ai-config loads the closure twice. ~/.claude/CLAUDE.md is a symlink to this repo's own CLAUDE.md, so it loads as global instructions while ./CLAUDE.md loads as project instructions --- byte-identical, counted separately. At today's figure that is roughly 608k tokens before a session has done anything, against a 200k budget. In any other repo only the single ~304k copy loads.