[Birfy]

Written-by: Dev

Fixes #895
Refs #796

A tree coordinator's system prompt carried Child load right now — Dev: working, Test: idle … (#507, #624). plugins/tree/index.js rewrites the tree block at every turn.start / turn.end of a tree Bot, so the coordinator's system prompt changed at every child turn boundary (and every minute for a stalled child). On DeepSeek DSH then appends the whole changed system prompt after the history (systemPromptUpdate: "in-history"); on other routes the cached head misses.

The load line is gone from the prompt (notesFor only: renames, roles and other fixed notes still refresh the block, and writeBlock skips an unchanged block). Coordinator rule 3 now says once, in fixed words: when more than one child fits a part, call list_children before sending it, and prefer an idle or working child to a busy one. Load only ever broke a tie between children whose scope fits, so a part with one fitting child costs no extra call. list_children already returns each child's load, openForMs and stalled live (#507, #624, #733), so the coordinator reads the same facts, fresher, when it routes. loadLine stays a pure helper in load.js (its own tests keep it); the prompt no longer calls it.

Token measurement (plugin test harness, same workload before/after)

Workload	main c90af02	this PR
3 children, one turn each (start + end)	root system prompt changed 6×	0×
Size of each change (root Chief, 5 children: identity section with the 3,100-char tree block)	≥ ~11,500 chars (~2.9k tokens) appended per change → up to ≈ 69k chars per delegation round	at most one list_children call when more than one child fits a part (a few hundred chars)

Verification at 3d4e268057209eda3ba85770b6a01abc095405a3 (rebased on b54f38a)

• TZ=UTC ARCH_DOCS_BASE=origin/main npm run verify → pass (core 1328/1328, plugins 395 + 125 + 125 + 332 + 37 + 16, arch-docs OK, patches check OK).
• Red at base: the updated Tree: show a coordinator its children's runtime load #507/Tree coordinator's system prompt changes at every child turn boundary (load line), so DSH re-sends it whole #895 test in plugins/tree/test/tree.test.js (prompt byte-stable across a child's turn start/end) fails on c90af02, passes here.

Not run: smoke bench, Desktop, a real model.

Docs: docs/subsystems/tree.md § Runtime load. Changelog: changelog.d/895-tree-load-line.fixed.md.

[Birfy]

Verdict: PASS at 3d4e268
Written-by: Reviewer

Ran:

• Fresh worktree detached at 3d4e268057209eda3ba85770b6a01abc095405a3, git status --porcelain empty, then TZ=UTC ARCH_DOCS_BASE=origin/main npm run verify → exit 0: core 1328/1328, plugins 395 + 125 + 125 + 332 + 37 + 16 all pass, arch-docs: OK (6 changed file(s)), ok core/ equals ds-bot/ + patches/ds-bot/
• Red at base: copied this tip's plugins/tree/test/tree.test.js onto c90af02, node --test --import ./test/resolve.mjs --test-name-pattern="895" test/tree.test.js → fail 1 (the prompt-byte-stable test), passes here
• git diff origin/main -- plugins/tree read in full; grep -rn loadLine → only load.js, its own tests and docs/subsystems/tree.md

Not run: smoke bench, Desktop, a real-model run.

Findings: no actionable findings.

• The coordinator still gets each child's load when it routes: list_children returns load, openForMs and stalled live, and its description names them (list_children does not describe the load/openForMs/stalled fields it returns #733); rule 3 tells the coordinator to call it when more than one child fits a part. Load only ever broke a tie between children whose scope fits, so a part with one fitting child costs no extra tool call.
• Renames and role edits still refresh the block on a turn boundary (refreshPrompt → writeBlock, which skips an unchanged block); with no load in it, a boundary that changed nothing else writes nothing (pinned by the Tree: show a coordinator its children's runtime load #507/Tree coordinator's system prompt changes at every child turn boundary (load line), so DSH re-sends it whole #895 test).

Security fence / display-name: not in scope (no remote text or fencing touched).

Disclosure: the same executor session that wrote this PR ran this review (owner's instruction to fix these together; Birfy is in selfReview). This is not merge authority by itself; the merge goes through enqueue-guard and the queue.