Coordination measurements

Collect selected frozen owner evidence or summarize a sanitized snapshot offline. This operation follows this reference; live team-round steps govern native work. It returns observational evidence and grants no task, release or production authority. Return the summary or receipt and finish this skill here. On a non-zero exit, report the diagnostic and finish here without starting live team work.

Collect selected evidence

Use the installed producer with explicitly selected frozen owner and supervision files and a fixed observation window:

HS=.tessl/plugins/jbaruch/herdr-policy/skills/herdr-foreman/; [ -d "$HS" ] || HS="$HOME/$HS"; [ -d "$HS" ] || case "$(git config --get remote.origin.url)" in git@github.com:jbaruch/herdr-policy|git@github.com:jbaruch/herdr-policy.git|https://github.com/jbaruch/herdr-policy|https://github.com/jbaruch/herdr-policy.git|ssh://git@github.com/jbaruch/herdr-policy|ssh://git@github.com/jbaruch/herdr-policy.git) HS=skills/herdr-foreman/ ;; *) echo "herdr-policy skill not found: install jbaruch/herdr-policy" >&2; exit 1 ;; esac
python3 "$HS/measurements.py" collect --owner /private/baseline/owner.json --supervision /private/baseline/supervision.json --start 2026-10-10T00:00:00Z --end 2026-10-10T01:00:00Z --output /private/baseline/capture.json

Omit --output for a bounded coordination_capture_summary (schema_version 1) containing summary schema 2, the sanitized snapshot digest and named collection exclusions. A large summary refuses without partial output; retain the full cohort and use a private report path.

With --output, the mode-0600 coordination_capture_report (schema_version 1) contains sanitized snapshot schema 2, private provenance schema_version 3 and summary schema 2. Its schema-1 coordination_capture_receipt gives path, digest, byte count, snapshot digest, summary version, cohort counts and exclusions. All three new record kinds set authorizes_progression: false. The herdr-foreman skill owns their shapes; the collector writes them and comparison readers read only supported versions. Existing reports and source evidence are never migrated. The snapshot can be passed to the existing calculator after extracting it from the private report.

The producer's EvidenceReader and export in skills/herdr-foreman/foreman/measurement_capture.py own source verification and cohort selection. Missing evidence remains unmeasured with collection exclusions. Integration, deployment, acceptance, direct foreman effort and activity samples remain unmeasured. Collection creates no recurring model turn, worker call, owner migration, state write or history ledger.

skills/herdr-foreman/foreman/measurement_capture.py owns MAX_OWNER_BYTES per owner/supervision file, MAX_NATIVE_BYTES per transcript, MAX_LEDGER_BYTES per captured prefix and MAX_EVIDENCE_BYTES across transcript/prefix reads in one invocation. MAX_SNAPSHOT_BYTES in the calculator bounds the sanitized export. Nonregular, linked, malformed or oversized owner inputs refuse. Native/ledger read failures leave named exclusions, including exhausted evidence budgets. Preserve originals and the full cohort; report the diagnostic or exclusions and finish here. Producer changes require a separately authorized implementation task.

Provenance retains normalized window, Python version, owner/supervision digests, actual producer digest, canonical reader file digests, verified evidence digests and exclusions. It contains no transcript bodies or private source-path map. scripts/export_owner_baseline.py uses this same producer and retains its explicit trusted-older-reader interface. Its exporter_sha256 identifies the relocated producer; its snapshot schema 2 and provenance schema_version 3 remain supported.

Summarize a prepared snapshot

HS=.tessl/plugins/jbaruch/herdr-policy/skills/herdr-foreman/; [ -d "$HS" ] || HS="$HOME/$HS"; [ -d "$HS" ] || case "$(git config --get remote.origin.url)" in git@github.com:jbaruch/herdr-policy|git@github.com:jbaruch/herdr-policy.git|https://github.com/jbaruch/herdr-policy|https://github.com/jbaruch/herdr-policy.git|ssh://git@github.com/jbaruch/herdr-policy|ssh://git@github.com/jbaruch/herdr-policy.git) HS=skills/herdr-foreman/ ;; *) echo "herdr-policy skill not found: install jbaruch/herdr-policy" >&2; exit 1 ;; esac
python3 "$HS/measurements.py" /private/baseline/snapshot.json

The command's stdout uses the shared foreman/context_output.py byte budget. For a complete report larger than that budget, retain the whole selected cohort and request one private output file:

HS=.tessl/plugins/jbaruch/herdr-policy/skills/herdr-foreman/; [ -d "$HS" ] || HS="$HOME/$HS"; [ -d "$HS" ] || case "$(git config --get remote.origin.url)" in git@github.com:jbaruch/herdr-policy|git@github.com:jbaruch/herdr-policy.git|https://github.com/jbaruch/herdr-policy|https://github.com/jbaruch/herdr-policy.git|ssh://git@github.com/jbaruch/herdr-policy|ssh://git@github.com/jbaruch/herdr-policy.git) HS=skills/herdr-foreman/ ;; *) echo "herdr-policy skill not found: install jbaruch/herdr-policy" >&2; exit 1 ;; esac
python3 "$HS/measurements.py" /private/baseline/snapshot.json --output /private/baseline/report.json

--output creates a mode-0600 report outside repositories and installed packages. Its file schema 1 contains the input byte digest, calculator digest, Python version and summary schema 2. Stdout returns a schema-1 receipt with path, digest, byte count and cohort counts. Identical repeats preserve the existing file; changed or foreign outputs refuse. An I/O failure may leave its newly created partial file; preserve that file and select a fresh output path after repairing the cause. Keep the finite comparison report and its source snapshot private while evaluating the preregistered cohort; publish supported aggregates and evidence links in GitHub. Once the comparison and evidence obligations are closed, retire its local artifacts. Reports never become a must-load history floor.

MAX_SNAPSHOT_BYTES in foreman/coordination_metrics.py owns the input budget. The command reads only the selected snapshot, performs no network or native team operations, and creates a file only for an explicit --output. Existing repo commands through scripts/coordination_baseline.py use this same owner.

Sanitized snapshot

The snapshot command reads only this file, emits one bounded summary or report receipt, and performs no network or team-state operations. A malformed export exits 2 with a diagnostic and no measurement. Keep snapshots and their raw-source maps outside the repository.

The exporter writes input schema 2. The analyzer also accepts schema 1 with missing endpoint times treated as unmeasured. Both have these required fields:

Optional task records:

• integration, deployment, acceptance: each has an at timestamp inside the observation window and an evidence array of opaque references. Both are required to count an evidenced stage. The exporter verifies the source; references alone do not independently prove an external effect. Omit events outside the window rather than silently counting preexisting outcomes.
• foreman: known active_seconds, exec_calls, output_bytes, and reconstruction_bytes. Each source event contributes once across the cohort; split a shared event by a recorded allocation or mark attribution unknown. A missing/null measurement remains unknown. Do not add overlapping activity.
• rework_rounds, escaped_defects: nonnegative measurements or null. An unmeasured defect rate is not zero defects.
• holds: entries with start, end (null when open), and reason from operator, policy, dependency, quota, outage, verification, unknown. Preserve operator/policy/dependency holds; the analyzer lists them without summing them into waits or subtracting them as if every held second were avoidable.

Optional assignment fields are native_started_at, native_completed_at, assessed_at, resolved_at, and report_bytes. Native timestamps require a verified enrollment/session match in the exporter. Retain unmatched records; do not infer native completion from a status banner or report modification time.

Schema 2 also accepts assessment_state and resolution_state: completed, pending, or unmeasured at the observation boundary. An explicit pending state requires source evidence of an open queue at that boundary. Missing time alone never proves pending work. A completed state without its measured time remains excluded from latency and exposure calculations.

The exporter marks canonical owner resolutions as completed or pending when their enrollment is available. Unavailable enrollments remain unmeasured. Assessment timing needs the verified ledger receipt; unavailable or changed receipts leave assessment unmeasured, including after owner resolution. The exporter cannot prove an open assessment queue from an absent receipt.

Before export, join immutable enrollment and native-session identities, retain the match decision in a private source map, fix the window, and record task selection/exclusions. Preserve missing evidence and changed/invalidated heads. Normalize clocks to offset-aware timestamps; never replace negative delays with zero. Remove transcript bodies, usernames, credentials, repository-local paths and production data; opaque ids preserve joins. Exporter/version and source digests belong in the private provenance record, not a public transcript dump.

Summary fields and interpretation

foreman/coordination_metrics.py owns derived summary schema 2. Readers select that supported version explicitly. Regenerate a summary from its sanitized snapshot; this operation never migrates or rewrites source evidence, owner state or previously saved reports.

latency_seconds reports native work, assessment and resolution distributions. Read each denominator and its named exclusions with that distribution; unknown or unusable endpoints remain visible. The calculations are owned by paired and distribution in the calculator. Assessment and resolution delays overlap and cannot be added. A queue delay alone does not establish avoidable effort.

waiting distinguishes verified pending queues from unmeasured endpoints. Each queue reports age statistics, an oldest list, omitted counts, known wait_task_seconds, covered tasks and exclusions. waiting and union_seconds in the calculator own the exposure calculations; OLDEST_PENDING_LIMIT owns the list budget. These are known waiting exposures across tasks, not foreman active effort. Read their coverage and exclusions together. Assessment and resolution totals overlap and cannot be added. Blocked work remains included. Holds retain their separate reason categories and never prove presumed waste.

activity describes possibly busy envelopes, not distinct physical workers. Duplicate ownership or a reused native session can overcount workers. This conservative proxy is not exact fleet utilization and must not be relabeled as the historical concurrency estimate without checking the private original sampling and identity-deduplication method. summarize owns its sampling calculation.

effort retains known and unknown measurement counts and reports output per evidenced accepted outcome when measurable. Its calculations live in totals and the effort projection in summarize. A null value is unmeasured, not zero. Failed and blocked work remains in the cohort. Compare bytes with bytes and time with time; no bytes-to-tokens or tokens-to-dollars conversion is assumed. GitHub integration, actual deployment and measured acceptance remain separate outcomes. An exporter assertion is distinct from verified external readback.