Rendered from docs/obligations/0198-nothing-states-how-much-of-a-session-s-budget-the-standing-instructions-consume-before-work-starts.md in the Headwater corpus. Every document on this half of the site is typed by the taxonomy the descriptor names: corpus.json.

Nothing states how much of a session's budget the standing instructions consume before work starts

Context

Every Claude Code session in this repository loads standing instructions before any task work begins. The load includes the global CLAUDE.md file from ~/.claude/. It also includes the project CLAUDE.md file at the repository root. The load adds system reminders, routing information from the Headwater governance engine, user context, and git status snapshots. This load is a fixture of every dispatch to an agent and every skill invocation.

The load consumes tokens from a session's budget. When this record was opened on 2026-09-16, it gave a range for the load that nobody had measured. No specification or planning rule stated the cost. No document stated what a session should assume or whether the cost matters.

Obligation

A measured baseline cost is what the corpus owes. A decision on its significance is also required. The instrument is either a quantified measurement or a decision record. A quantified measurement states the token count under defined conditions. These conditions include agent type, codebase size, and session mode. A decision record rules measurement out of scope and states what a session should assume.

Discharge

Discharge is satisfied by one of two paths.

First path: a measurement taken over a representative set of sessions. These sessions can be drawn from the test grid or from real sessions. The cost must be recorded in a specification, tutorial, or decision document. That document must state the conditions under which the cost holds.

Second path: a decision record. This record rules measurement out of scope. It states explicitly what an agent or planner should assume when allocating budget.

In either case, the document must be retrievable from this obligations register via a relation. A session or agent reading the corpus for budget guidance must be able to find it.

How the first path paid this record

The change for #1473 took the first path. Spec 5 states the figure and its conditions, and this record reaches it through its traces_to relation.

The source is the batch of #1384, recorded in PR #1465. It ran on 2026-09-30 at the pin 109abba7, with claude-sonnet-5, seed 980 and a turn cap of 80, on one host. The batch has one log for each of its 660 sessions. The table below uses the 540 logs of the campaign tier.

The instrument is the prompt before the first turn. The first assistant event of a log carries it as input_tokens + cache_creation_input_tokens + cache_read_input_tokens. The harness repeats that usage on each streamed part of one message, so the first line is the whole prompt. For each probe, the median of the present arm minus the median of the absent arm is the difference.

The absent arm removes CLAUDE.md, .claude/, .githooks/ and .headwater/. In the init event of each log, the present arm shows 42 skills and 14 agents. The absent arm shows 33 skills and 5 agents. The present arm lists 12 more slash commands: the 9 skills and the 3 commands next, next-run and product-owner. So the difference is the project CLAUDE.md and the output of the intent hook on the first prompt. It also holds the descriptions of 9 skills, 3 commands and 9 agents. The log does not divide the difference among these five parts.

This command, run as written, printed the table that follows it:

cd /home/james/.claude/jobs/1338ac1e/tmp/issue-1384/full/campaign/sessions &&
for d in L[1-6]-campaign-*; do
  printf '%s %s\n' "${d%-r*}" "$(jq -r 'select(.type=="assistant") | .message.usage | .input_tokens + .cache_creation_input_tokens + .cache_read_input_tokens' "$d/raw.jsonl" | head -1)"
done | sort -k1,1 -k2,2n | awk '
  { k = $1; v[k, ++n[k]] = $2; if ($2 > hi[k]) { hi[k] = $2; c[k] = 0 } if ($2 == hi[k]) c[k]++ }
  function med(k) { return (n[k] % 2) ? v[k, (n[k] + 1) / 2] : (v[k, n[k] / 2] + v[k, n[k] / 2 + 1]) / 2 }
  END {
    print "| probe | present median | absent median | difference | present sessions at the higher value | absent sessions at the higher value |"
    print "|---|---|---|---|---|---|"
    split("L1:L2:sufficiency:4 L3:L4:navigability:3 L5:L6:discovery:2", pairs, " ")
    for (i = 1; i <= 3; i++) {
      split(pairs[i], f, ":")
      for (p = 1; p <= f[4]; p++) {
        a = f[1] "-campaign-present-p" p; b = f[2] "-campaign-absent-p" p
        printf "| %s p%d | %d | %d | %d | %d of %d | %d of %d |\n", f[3], p, med(a), med(b), med(a) - med(b), c[a], n[a], c[b], n[b]
      }
    }
  }'

probe	present median	absent median	difference	present sessions at the higher value	absent sessions at the higher value
sufficiency p1	43007	35586	7421	29 of 30	24 of 30
sufficiency p2	42780	35530	7250	28 of 30	26 of 30
sufficiency p3	42903	35562	7341	25 of 30	27 of 30
sufficiency p4	42764	35499	7265	28 of 30	22 of 30
navigability p1	42487	35441	7046	26 of 30	26 of 30
navigability p2	42521	35448	7073	27 of 30	23 of 30
navigability p3	42711	35496	7215	29 of 30	23 of 30
discovery p1	42752	35527	7225	26 of 30	21 of 30
discovery p2	42399	35416	6983	29 of 30	25 of 30

Over the 9 probes, the median difference is 7,225 tokens, and the range is 6,983 to 7,421. The absent medians are 35,416 to 35,586, so the difference adds about 20% to the absent baseline.

Each probe takes a higher value and, in some sessions, a value 2,316 tokens lower. The cause is the host and not the tree. Over all 660 sessions of the batch, the init event shows the state of each connector. In all 100 sessions that started before the claude.ai Claude Docs connector was connected, the prompt was lower. In 559 of the 560 sessions that started with it connected, the prompt was at the higher value. The one exception started without a different connector and fell by 194 tokens. The lower value occurs in both arms, so the median of each arm stays at the higher value. The documentation absent arm gave the same medians as the campaign absent arm. So the removal of docs/ does not change the prompt before the first turn.

The logs are in a job directory outside the tree, and the tree keeps only this table. The figure is for the tree at 109abba7. CLAUDE.md changed after that tree, so the figure is a dated measurement.