Weekly team metrics¶

The team's weekly scorecard, defined once here and computed by one SQL file over the data plane. hangar's standup line and Slack digest (hangar#168) cite this page and read latest.json, never computing their own.

The only implementation is src/flightdeck/metrics/weekly.sql: one DuckDB view per metric over the data plane, with each numerator and denominator repeated in a header comment. A change to a definition changes this page and that file in the same PR.

flightdeck metrics weekly [--week 2026-W45] [--format md|json|slack] [--person P] [--repo owner/repo]

It reads Parquet only (FLIGHTDECK_DATA_URI), so it runs with hangar powered off. With FLIGHTDECK_REPORTS_URI set and no --person or --repo, it also writes <reports>/weekly/<week>.json, <week>.md and latest.json (flightdeck.weekly/v1). The nightly compactor job refreshes them for the last complete week.

Conventions¶

Week¶

ISO week, Monday 00:00 to the next Monday 00:00 UTC. An event belongs to the week its timestamp falls in.

Agent job¶

A hangar job with source = job: delegated work, from the API, GitHub, Slack or Jira. Experiment trials and shadow runs are not delegation and are left out. An agent PR is a PR a hangar job opened (pr.open job event).

Terminal¶

A job is terminal at the first of: its PR merged (pr.merge), its PR closed unmerged (pr.close), or it was abandoned (abandon, that is, cancelled). A job with no PR is terminal when its last run's job_state is succeeded, failed or cancelled.

Counted¶

A terminal job counts when it ran at least one turn and its last run's outcome_class is not infra_error (hangar#166). A job cancelled before it started does not count. Infra-error runs are reported as a count next to the metrics, never silently dropped.

Intervals¶

Rates come with a Wilson 95% interval (flightdeck.stats.core.wilson), always printed with k of n, because six people give small numbers. The markdown shows the week next to the three weeks before it.

Scope¶

--person and --repo narrow every job-based metric. Per-person numbers are for that person and admins only (hangar's owner scoping): printed, never written to the shared report files. The Slack team view is team-wide. Agent share is team or repo scope only.

The metrics¶

Metric	Target	View
Agent PR yield	≥ 60%	`weekly_yield`
Agent share of merged PRs		`weekly_agent_share`
Review wait	p50 < 4 h, p90 < 24 h	`weekly_review_wait`
Green to merge		`weekly_green_to_merge`
Human steers per merged PR		`weekly_steer_rate`
Abandon rate	< 25%, with a reason	`weekly_abandon`, `weekly_abandon_reasons`
14-day revert rate		`weekly_revert`
Code survival at merge		`weekly_survival_at_merge`, `pr_outcomes`
Human-edit share		`weekly_human_edit_share`, `pr_outcomes`
Cost per merged PR		`weekly_cost_per_merged_pr`
Active delegators	6 of 6	`weekly_active_delegators`

Yield¶

Agent jobs whose PR merged ÷ counted agent jobs that became terminal in the week. A rate, with its interval.

PRs merged in the week that a hangar job opened ÷ all PRs merged in the week on the team's repositories. The denominator is the prs table, from GitHub, humans' and bots' PRs included. An agent PR is matched by its URL.

Review wait¶

Hours from the agent PR opening (pr.open) to its first counting review (review.first: the first approval hangar's landing counts, bots excluded), for agent PRs first reviewed in the week; p50 and p90. Until hangar records the time a draft is marked ready (hangar#168 ready_at), a PR opened as a draft includes its draft time.

Green to merge¶

Hours from the PR's checks going green to its merge (hangar's green_to_merge_seconds on pr.merge), for agent PRs merged in the week; p50 and p90.

Steer rate¶

Human touches on the agent jobs merged in the week ÷ those jobs. A touch is a steer or interrupt during a turn (run.json steers.n), or a question the agent asked that a person answered (questions). A follow-up turn, every job turn after the first, is a touch too. Corrections-only counts follow once hangar classifies steers (hangar#162).

Abandon¶

Counted agent jobs terminal in the week that were abandoned or whose PR closed unmerged ÷ the same denominator as yield. Broken down by hangar's abandon reason: wrong_approach, bad_ticket, flaky_env, too_big, superseded, other. A closed PR or cancelled job without one is unlabelled.

Revert¶

Take the agent PRs merged two weeks earlier. The rate is those a PR merged within 14 days of them reverts ÷ all of them. The lag makes every PR in the cohort old enough for its 14 days.

A revert is a PR whose body is GitHub's revert-button text (Reverts owner/repo#N), says This reverts commit <sha> for the PR's merge commit, or is titled Revert "<title>". Reverts pushed straight to the branch are not seen until hangar emits revert job events.

Survival at merge¶

Defined in Outcome labels.

Cost per merged PR¶

The summed cost_usd of every run of the agent jobs merged in the week ÷ those jobs. Review jobs (hangar#161) are separate jobs, and not included until they link to the job they review.

Active delegators¶

People with at least 3 agent PRs merged in the week, out of the people with an agent job run that week.

Infra errors¶

Runs in the week whose outcome_class is infra_error: left out of yield and abandon, and reported as a count so the exclusion is visible.

Where the data comes from¶

The decision recorded for #6: facts hangar knows come from hangar, facts only GitHub knows come from GitHub.

Data	Source	Table
Jobs, runs, cost, steers, outcome class	hangar's outbox `run.json`	`runs`
PR opened, first review, merged (with green-to-merge and outcome labels), closed, abandoned (with reason)	hangar's outbox job events	`job_events`
All merged PRs on the team's repos, reverts	GitHub REST, nightly (`flightdeck ingest github`)	`prs`

The issue's alternative, hangar emitting review and revert job events, is half taken: hangar emits review.first and revert, but the revert view still reads prs.