Herdr team coordination policy for AI coding agents
93%
Does it follow best practices?
Run evals on this skill
Adds up to 20 points to the overall score
Low
Low-risk findings worth noting
Read the co-shipped skills/herdr-foreman/references/worker-context.md and the selected general policy, then
perform only the current admitted assignment and selected responsibilities.
The executor owns progression and local teardown. Review and testing are
read-only. A release worker follows its complete selected release skill and
binding references, preserves the writer checkout and repairs no source.
Do not read the foreman workflow or reconstruct prior rounds.
A report is the enrolled file delivered by the exact native assignment. The
final chat line is exactly REPORT: <absolute path>. File existence, pane text,
an exit code or a report's own claim is insufficient delivery proof. Native
launch, permissions, model/process, composer and final-source checks remain
binding. Their canonical assignment record owns tier provenance; repeating
that metadata in prose grants no extra authority.
Every report contains exactly one operative WORKFLOW: <JSON object> line,
outside examples, quotations, indentation and code fences. A reviewer/tester
also carries exactly one operative VERDICT: approved|blocking line. The
claim must match that verdict. The complete schema is validated by
skills/herdr-foreman/foreman/workflow_contract.py; unsupported, missing, duplicate or conflicting
claims do not advance the task.
The schema-1 claim fields are schema_version, task, head_revision,
responsibilities, outcome, acceptance, checks, findings, correction,
pull_request, review_snapshot, review_evidence and publication.
implementation,
review, tests or release. A bounded verifier supplies both review and
tests.responsibility, name, result and evidence. Results
are passed, failed or unavailable. A successful outcome needs actual
successful evidence for every selected responsibility.criterion (one-based number), state (met or unmet) and evidence.
Other responsibilities set acceptance to null. A verified outcome requires
every criterion met and every claimed check passed.built or blocked; verification outcomes
are verified, blocking or blocked. Release outcomes are delivered,
blocking or blocked; ready records preparation only and cannot complete
delivery.id, severity (blocking, advisory or
uncertain) and status (open, resolved or disputed). Review/test and
blocking release outcomes supply explicit correction facts:
scope_unchanged, cause_settled, contradictory_evidence, repeated_defect,
non_converged, changed_production_plan. Other outcomes use an empty
findings list and null correction. The executor derives current authority
and evidence facts itself.The publication claim has schema 1, baseline, baseline_at, publisher_run,
tag and tessl_version. Report actual canonical-helper results. A Tessl
baseline/time must come from before the actual merge; preserve unpublished
only when its helper verified that sentinel. Nonselected channel fields are
null. No publication uses a null run id; a task without publications uses null
publication identities. The initial adapter supports the admitted
merge-triggered Tessl/ACR publisher, bound to workflow, merge commit, push and
base ref. Other publication/deployment adapters require their actual owner.
Preserve current report gates and operator holds before merge or publication. Return blocking source corrections to implementation. Required publication, deployment and acceptance need their actual evidence; merge alone is not substituted. The role template supplies the current release procedure.
.tessl-plugin
rules
skills
herdr-foreman
classify
foreman
references
specialists
templates
herdr-observer
herdr-optimizer
herdr-standup