Skip to content

analysis: failures name their stage - #171

Merged
trick77 merged 6 commits into
masterfrom
analysis-run-stage
Nov 19, 2026
Merged

trick77 merged 6 commits into
masterfrom
analysis-run-stage

Conversation

@trick77

@trick77 trick77 commented Nov 19, 2026 •

Copy link
Copy Markdown
Owner

fail logged stage=llm for every failure after the claim. It now takes the stage: marshal_input, queue, llm, validate, marshal_output, finish (FinishAnalysis failed); store stays for FailAnalysis failing inside fail. acquire hands back a slot won on a dead context, since select picks either when both are ready. A validate failure names the finish reason as finish_reason=, only when there is one. Field names unchanged.

@trick77 trick77 changed the title analysis: only manual runs check unchanged input; failures name their stage analysis: failures name their stage Nov 19, 2026
@trick77
trick77 force-pushed the analysis-run-stage branch 2 times, most recently from 90f78f0 to 5734811 Compare November 19, 2026 08:36
Deps.TeamNames and the teams handler sorted the configured teams themselves; config.Config.TeamNames already does.
Compute, Assess and build defaulted a nil previous dataset or baseline;
production always passes a loaded previous period (api read, analysis
assess). The default moved into the tests' noPrevious and buildAlone;
behaviour unchanged.
stageHistogram and histogram share edgeAbove. ClosedPeriod, ClosedPrevious and PreviousPeriod are whole weeks ending at a Monday: weeksBefore takes that end. NewPeriod ends at now, mid-week, so it keeps periodFrom.
due and MaxAge encoded each window's schedule twice (12: every second week / 2, 26: first Monday / 5). Both read cadenceOf now; an unnamed window stays weekly, max age 1. A test holds every cadence's widest gap to its maxAge; NextRun searches at most 2*maxAge+1 weeks, so a cadence never due cannot spin.
run looked up the last done analysis on every trigger but used it only for the manual unchanged-input skip; scheduled runs no longer query it. Run's doc said any unchanged input is not asked again; only a manual run's is. DoneAnalyses breaks a finished_at tie by id, as DoneWeeksBefore does.
fail logged stage=llm for every failure after the claim. It now takes the stage: marshal_input, queue, llm, validate, marshal_output, finish (FinishAnalysis failed); store stays for FailAnalysis failing inside fail. acquire hands back a slot won on a dead context, since select picks either when both are ready. A validate failure names the finish reason as finish_reason=<x>, only when there is one.
@trick77
trick77 merged commit f81b667 into master Nov 19, 2026
1 check passed
@trick77
trick77 deleted the analysis-run-stage branch November 19, 2026 08:44
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant