plat 136

← Pulse platform issue index

PLAT-136 — every run in the schedule popup showed the same cost, because every run claimed the same folder

Coordination	Value
Assigned agent	unassigned
Ticket state	implemented — rotation now repoints run history; live reverify pending
Last synchronized	2026-08-18

• Priority: P1 — the run history is the surface used to judge whether a scheduled workflow is healthy and what it costs. Every row reported the same spend, so it could not answer either question.
• Owner: pkg/orchestrator/agents/workflow/step_based_workflow/controller_run_manager.go (rotatePairedIterationZero).

How it surfaced

The schedule popup for hetzner-ssh showed seven runs with wildly different durations — 5m 44s, 43m 14s, 1h 19m, 1h 39m, 23m 57s — and an identical $255.48 / 300.92Mt / [1g] on every row.

Root cause

Every run executes in runs/iteration-0, and its history entry records exactly that. But iteration-0 is only the live slot: the next run rotates it to a permanent iteration-N (rotatePairedIterationZero). Nothing updated the history entry when that happened.

Measured on hetznerssh: 24 of 25 entries recorded run_folder: iteration-0 while the folders on disk were iteration-0, 21, 22, 23, 24, 25. One entry's own error text names iteration-25 while the entry itself says iteration-0 — the contradiction is visible inside a single record.

The popup looks cost and tokens up per run folder. Rotation already archives the cost records (ArchiveRunCostPaths) and the evaluation scores (archiveEvaluationScoreRunFolder) to the new name — both correct. So every history row resolved to the one folder still called iteration-0 and rendered the current run's spend. The rows were not stale; they were all reading the same live cell.

Run history was simply the third thing that names the folder and the only one rotation never repointed.

Fix

ArchiveScheduleRunFolder, called from rotatePairedIterationZero alongside the two archive calls that already existed, repoints the history entry that owned iteration-0 at its new permanent name.

Only the newest finished claimant is repointed. Runs still holding the live slot are skipped — a running run is obviously in it, and a capacity-suspended run resumes into the same folder when the provider window reopens (PLAT-101), so repointing it would send the resume looking for evidence under a name it never wrote to.

The file is decoded as generic maps rather than a typed struct: the typed entry lives in cmd/server, which this package cannot import, and a local struct would silently drop every field it does not know about on write.

What this does not fix

Historical entries stay wrong. The runs that produced iteration-21..24 recorded iteration-0, and the mapping from those entries to those folders is not recoverable — the only evidence would have been the rotation that already happened. Guessing would replace a visible wrong answer with an invisible one.

Rows for runs from before this change will therefore keep showing the live slot's cost. Rows from here on will show their own.

Acceptance

• After two scheduled runs, the older run's history entry names the rotated iteration-N folder and the newer one names iteration-0.
• The schedule popup shows different cost/token totals per row.
• A capacity-suspended run's entry still names iteration-0 after an unrelated rotation, and resumes into that folder correctly.