[Birfy]

Written-by: Dev4

Reproduce

turn.end from the core feed carries the turn it is about (turnId, e.g. <session>#2), and plugins/tree/index.js handed only botId to the tracker. The tracker kept bare start times and closed whichever turn was oldest, so a later turn ending first left the wrong age behind — the very state #624 exists to expose:

turn.start  s#1  at T+0
turn.start  s#2  at T+10min
turn.end    s#2  at T+12min      # the newer turn ends
list_children -> { load: 1, openForMs: 120000 }   # 2 min: s#2's age, though s#1 has been open 12

Two Bots' turns were merged into an age that belongs to neither, and a coordinator reading stalled could miss a genuinely stuck turn behind a busy one.

What changed

• plugins/tree/src/load.js — each open turn is kept as { turnId?, at } instead of a start time, and end(id, turnId) closes that turn. No match — an id never seen, or none at all — still drops exactly one turn, the oldest, so the count cannot drift and an older caller keeps working. of, openFor, snapshot and loadLine are unchanged in shape.
• plugins/tree/src/tree.js — noteTurnStart / noteTurnEnd take the turn id and pass it on.
• plugins/tree/index.js — both feed handlers hand e?.turnId over instead of dropping it, with the reason written where the wiring is.
• plugins/tree/test/load.test.js — #696 case: two turns open, the later one ends, and the age left is the older turn's; an unknown id and a missing id each drop exactly one turn (the oldest), and a named end closes the turn it names.
• plugins/tree/test/tree.test.js — the same at feed level through list_children and the coordinator's stored prompt: after turn.end for s#2 the child still reads working (12 min, no turn ended), not 2 minutes.
• docs/subsystems/tree.md — the #624 section now says an end closes its own turn, and names the fallback.

Still a routing hint, never a gate: nothing here refuses, ends or reassigns work.

What I ran

Tip SHA: 10e87463cb16f7e335d926fe04b5db9f30262c38 (Node 22.23.3, TZ=UTC, base 2f538019, run in a checkout of that base with these six files applied):

cd plugins/tree && node --import ./test/resolve.mjs --test --test-concurrency=2 "test/*.test.js"
  -> # tests 102 / # pass 102 / # fail 0
cd plugins/tree && node --import ./test/resolve.mjs --test test/load.test.js test/tree.test.js
  -> 91/91
node scripts/lint-plugins.mjs          -> 0 errors (pre-existing warnings only), @flybot/tree packs and imports
node core/scripts/check-arch-docs.mjs  -> arch-docs: OK (6 changed file(s))  [merge base = the PR base]
node .github/scripts/changelog.mjs check -> 13 fragment(s) in changelog.d/ are well formed

Red without the source change

With end() back to turns.shift() — the turn id ignored, exactly as on main — and both tests and call sites left alone:

not ok 5  - #696: an end closes its own turn, not just the oldest open one
  expected: 720000   actual: 120000
not ok 82 - #696: two turns open, the later one ends, and the age left is the older turn's
  expected: 720000   actual: 120000
# tests 91 / # pass 89 / # fail 2

120000 is s#2's 2 minutes where s#1's 12 should be — the defect itself. Restoring load.js gives 91/91, then 102/102 for the whole plugin.

Not run

• npm run verify end to end: this box hits spawnSync /bin/sh EAGAIN at the default concurrency, so the steps above were run one by one.
• The core suite: this change touches no core/ file and no host wiring, only the plugin's own tracker and its two feed handlers; the plugin suite and CI's tests-ok cover it.
• No new end-to-end bench of a two-turn Bot: the two cases drive the same feed events the host emits (turn.start/turn.end with turnId), asserted through list_children and the stored prompt.

Risk

T2 — plugins/tree/** plus docs/subsystems/tree.md, 6 files, +87/−26. No core/ path, no .github/**, no security surface. An end with no id behaves exactly as it did before this PR, so no caller that omits the turn id changes behaviour.

Fixes #696

[Birfy]

Verdict: PASS at 10e8746
Written-by: Reviewer

The defect is real and the fix closes it at the right layer. index.js now hands the feed's turnId to the tracker, load.js keeps { turnId, at } per open turn, and end() splices the turn it names instead of always the head of the array. Because pushes append and splices keep order, openFor() still answers the oldest turn that is really still open, and of/snapshot/loadLine keep their shape, so no reader changes.

The fallback is the part I checked hardest: end() with no id, or with an id it never saw, still removes exactly one turn — the oldest — so the count cannot drift and an older caller that passes only botId behaves as before. Two properties make that safe. Two turn.ends for one turn cannot happen: flybot-feed.js:252 fires only while entry.turn !== undefined and clears it, and a turn/start that closes a crashed turn emits its one end first (:203-215) — so the "unknown id" branch is only reached by a turn.start the plugin never saw (a plugin loaded mid-turn, or wave-off), and it stays inside the Bot's own list.

Ran: cd plugins/tree && node --import ./test/resolve.mjs --test --test-concurrency=2 "test/*.test.js" (TZ=UTC DSH_HARNESS=/opt/flybot)

• tip 10e87463: 102 tests, 102 pass, 0 fail.

Ran: GET /repos/ThinkFlowLab/FlyBot/commits/10e87463/check-runs

• 9 checks, all completed/success; mergeable: true, mergeable_state: clean; 6 files, +87/−26, base 2f538019.

Ran (red without the source change): copied the tip, restored plugins/tree/src/load.js, plugins/tree/src/tree.js and plugins/tree/index.js to 2f538019, kept both test files, same command

• 102 tests, 100 pass, 2 fail — exactly #696: an end closes its own turn, not just the oldest open one and #696: two turns open, the later one ends, and the age left is the older turn's, the second reading the defect itself (120000 where 720000 belongs).

Not run:

• npm run verify end to end (this box runs out of processes at its default concurrency); the pieces above were run one at a time, and the package job covers the rest at this tip.
• No two-turn Bot end to end: the two cases drive the same turn.start/turn.end events the host emits, read back through list_children and the coordinator's stored prompt, which is the surface a coordinator actually reads.

Not blocking: the doc line now says "an id never seen, or none at all" drops the oldest, which is accurate but reads as normal traffic. The feed cannot produce that case for a turn the tracker saw (flybot-feed.js:252-259), so it is worth one clause — "a start the plugin missed" — because otherwise a later reader will look for a duplicate-end bug that cannot happen and miss the one that can (a plugin loaded while a turn was already running).