TASK-135 — an existing row can change track

[ranjiao]

The gap

--track was accepted by exactly two commands: add, at creation, and route, which turns an intake row into a task. There was no perry-task track <id>, and status --track is refused.

So a project that declared a second track started it empty and could not move any existing work onto it. This repository declared intake (mode queue) on 2026-08-20 and it is still empty, while six rows that genuinely arrived rather than being decomposed belong on it and could not get there. A track with no rows is not a mode running on a live track — which is why the declaration did not meet KR-O1.1.

Source: perry/evidence/2026-08/TASK-133-track-experiment.md, spec perry/evidence/2026-08/TASK-135-spec.md.

What ships

perry-task track <ID> --track <name> [--stage S] [--arrived YYYY-MM-DD] [--reason "…"]

Onto a queue track it stamps what route stamps — Arrived, and the track's first post-intake Stage via the same entry_stage that add and route call. today − Arrived is the number every SLA check measures, so a second entrance that omitted it would make Arrived mean one thing per entrance and silently exempt every moved row from the only clock governing it.

An existing Arrived is carried, never restamped. A queue → queue move that reset the clock to today would erase an in-flight breach — the same exemption, arriving through the other door. --arrived overrides.

Off a track that reads neither, Stage / Stage since / Arrived are cleared, and what they held is recorded in the journal line and the event (stage_from, arrived_from). This is deliverable 2's decision, and it is drop's precedent applied one field down. Arrived is not provenance, it is a queue's clock, and one left on a row no queue governs is a live-looking number the next SLA reader picks up. It is already load-bearing in the negative direction and measured: cmd_route had to stop writing Arrived onto pipeline rows precisely because a non-empty arrived hides the row from rows_with_no_computable_age — the finding that says "this row has no clock at all". Keeping it would suppress the one report that could notice the move. The journal is append-only, so the cell goes and the fact does not.

An undeclared track is refused by name, in check_role's shape:

perry-task: refused — track 'ops' is not declared in `.perry/config.md § Tracks`.
Declared: main, intake, press. Add a row to that table naming its mode, or name
one of those. Nothing was written

It used to print a Python list repr and stop, which names the set and not the way forward. Nothing is created — a tool that invented the track would turn a typo into a second track every count then reports as real.

The two seams that would have half-landed the write

• commit's changed map carries ("track", "stage", "stage_since", "arrived"). Listing only track would land the new track in the store beside yesterday's clock; the board is rendered from the store, so the move would half-revert on the next render. Asserted against perry-tasks render --write.
• widen_row_section asks the widening question for both section shapes. Board.find returns P0/P1/P2 or a project's own heading, and the two widenings live on different methods — cmd_stage widens only in the first case, so on the board shape --group exists for a stamp is dropped silently at exit 0. A track move writes four such cells.

Verification — the spec's six items

1. Six rows move. TestSixRowsMove — six existing rows reach a declared queue track, the board renders Arrived and Stage, and perry-state --json reports stage_counts: {"triaged": 6} where it reported {} before.
2. The refusal is by name. TestTheRefusalIsByName — names the track, lists the declared ones, exit 1 with no traceback, creates no track and writes no event.
3. Both directions of the field question. TestBothDirectionsOfTheFieldQuestion — onto a queue track, then off it, asserting the store record both ways, plus the row that already had an Arrived before the move.
4. An event per move. TestAnEventPerMove — one track event per move, field: "track", reaching both timeline[] and perry-task events.
5. No drift. TestTheProjectionAgreesAfterAMove — perry-lint --json reports no store-drift finding after a move, and after a move that clears the clock, which is the direction a renderer silently skips.
6. python3 tests/parallel -j 4, bash tests/run, python3 bin/perry-lint, git diff --check — all below.

Tests

	baseline (6893da6)	after
python3 tests/parallel -j 4	65 modules · 1930 tests · 1 red	66 modules · 1960 tests · 1 red

The one red is test_diagnose with user_load.dangling non-empty — TASK-126, fixed on unmerged PR #22, inherited by every worktree cut from this branch. Identical before and after; nothing else is red. bash tests/run agrees. perry-lint: 0 errors, 3 pre-existing NS-01 warnings, 0 rows drifted. git diff --check clean.

Out of scope, and honoured

git diff -- perry/ is empty. The six live rows are not moved — this ships the tool; using it on this project's board is a separate act the user takes.

Note on the file count

feat/work-modes is ahead of origin, so this PR carries those commits until the branch is pushed. This row's own code diff is 6 files, +713 / −18: bin/perry-task, tests/test_track_move.py (new, 30 tests), tests/test_prioritize.py, schema/task-list-contract.md, schema/events-list-contract.md, work/reference/subcommands.md.

The contract-doc edits were not optional: EVENT_FIELD gains a key, and tests/test_prioritize.py binds that map's size to a number written in the contract's prose. That guard hardcoded of the fourteen in its own regex, so the day a fifteenth task event landed the doc could be corrected and the test would stop matching altogether and pass by finding nothing — it now reads both numbers out of the prose and checks the total against TASK_EVENTS.

🤖 Generated with Claude Code