# Session Coordination > Coordinate delegated agent sessions with bounded assignments, asynchronous background scheduling, soft concurrency caps, context-window rotation, independent review cycles, and logged cleanup across native child sessions and Herdr-managed agents. Use when managing multiple agents, panes, tabs, branches, stages, or long-running tasks while keeping the coordinating session small. - Skill: `timmo001/session-coordination` (Agent Skill) - Install (CLI): `npx skillmds@latest add timmo001/session-coordination` - Raw SKILL.md: https://api.skillmd.com/api/skills/timmo001/session-coordination/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: AI & ML - License: Apache-2.0 - Author: timmo001 (https://skillmd.com/u/timmo001) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/timmo001/session-coordination --- # Session Coordination Manage scope, dependencies, ownership, approvals, blockers, and completion evidence. Delegate substantial investigation, implementation, execution, verification, review, and large reads. ## Route To Owning Skills - Load `changeset-scope` before bounded delegation and pass the same boundary to every child. Surrounding reads are context only. - Load `staged-implementation` when work has independently reviewable stages. - Load `herdr` before Herdr control and `herdr-workflows` for session recovery or worktree transfer. Those skills own commands, topology, lifecycle, targeting, safety, and cleanup mechanics. - Load `git-context` for Git state, `gh-stack` for stacks, and `git-commit` only after explicit commit or push authorisation. Those skills own Git operations and approval details. - Require review sessions to load `code-review`, `changeset-scope`, and matching specialist skills. - Tell each child which repository and domain skills apply. Do not load its implementation skills into the coordinator's context. ## Assign And Schedule 1. Give every assignment an owner, dependencies, acceptance criteria, permitted mutations, verification, and compact return format. Choose the smallest capable set of sessions. 2. Use a soft cap of four actively working delegated sessions and three coordinator-created background panes in the current Herdr tab. Queue excess work. Exceed either cap only with user approval, and never create topology merely to bypass a cap. 3. Run independent or long-running work asynchronously. Native background sessions are only for bounded baseline research. When coordinating inside Herdr, use visible Herdr-managed sessions for implementation, execution, verification, and review; keep their panes unfocused while they work. Advance other ready work and join results only when a dependency needs them. Do not poll or duplicate running work. 4. Keep conflicting work sequential unless the user approves separate topology. Reuse a session only for the same assignment, its review fixes, or unfinished verification. New scope or a new independently reviewable outcome gets a fresh session. 5. Use a fresh read-only session for every review cycle. Route verified findings to the owning implementation session, then use another fresh reviewer. 6. Honour an explicitly requested agent runtime. Keep runtime selection separate from the agent profile selected inside that runtime: never convert a runtime name into a native profile argument such as OpenCode's `--agent`. Herdr's `agent start --kind` selects an integration's default executable; it is not a runtime-version selector. For a Herdr-managed alternate runtime, treat the configured launcher path as authoritative. Verify that exact path in the coordinator before creating its pane; never resolve a bare runtime command from `PATH` or use the target pane for launcher discovery. Start the exact launcher with `herdr pane run`, then use `herdr pane process-info` to confirm its foreground `argv` matches the launcher or its documented exec target before prompting the session. When the user does not choose a runtime, match the coordinator's runtime. Outside Herdr, use host-native child sessions. Inside Herdr, use Herdr-managed sessions except for the bounded baseline research allowed above. In OpenCode, keep V2 parents on V2 children and V1 parents on V1 children. Use Pi or another Herdr-supported agent kind only when the user requests it. The assignment phase is complete when every ready item is either owned, queued by the cap, or blocked on a named dependency or user decision. ## Control Context - Read only short briefs and targeted slices directly. Delegate full diffs, transcripts, logs, generated output, long history, and broad searches. Keep direct `context_git_context` calls free of `diff` and `branchDiff` payloads. - Every delegated brief must require compact results and compliance with the active context warning system. At a warning, stop adding scope and rotate after the smallest safe unit. At critical, stop unless that would leave an unsafe transition incomplete. - A successor receives only current scope, decisions, state, blockers, verification, and direct native or artefact references. Do not replay the old transcript. - Track role, runtime, native handle, scope, status, and blocker. Use the host's native session ID when available and Herdr agent plus pane IDs for Herdr-managed sessions. Context control is complete when each active session remains below its warning band or has a recorded successor rotation in progress. ## Record And Clean Up - Name coordinator-created Herdr agents with `coord-`. - Before closing any owned session, write and read back `~/.cache/agent-coordinator/sessions/-.md`. Include its handle, runtime, role, scope, repository or worktree, decisions, result, changed paths or findings, verification, blockers, artefact references, and closure reason. - Store no transcripts, large output, secrets, notes, or session handoff documents. Session records are lifecycle indexes, not durable project plans; other loaded skills still own project handoffs or notes when those are needed. - Follow `herdr` ownership and state safeguards when closing or recovering `coord-` sessions. If the record cannot be produced or verified, leave the session open. Host-native sessions remain referenced by their native ID. - Check for stale owned sessions at the start and end of a coordination run. Close only settled coordinator-owned topology with a verified record. - Treat lifecycle records as temporary cleanup guards. After every session from the current run is settled and closed and its result is included in the final summary, remove that run's record files one at a time and verify each path is gone. Never bulk-delete the directory or remove records owned by another concurrent coordinator run. Cleanup is complete when no settled coordinator-owned pane, tab, workspace, or agent remains open and no lifecycle record created by the current run remains. ## Approvals And Delivery - Use direct read-only Git and GitHub commands only for concise management facts; delegate detailed judgement. Delegate mutating Git and stack work to the owning implementation session. - The coordinator owns the workflow decision for consequential mutations. Pass the exact approved operation and scope to the child. Native children retain their own permissions, so runtime permission prompts may still appear. - Resolve conflicting conclusions and failed verification rather than forwarding them without analysis. If a child fails, clarify or reassign the work instead of absorbing its operational task. - Report completion only when implementation, independent review, and required verification satisfy the acceptance criteria. State skipped work and pending approvals explicitly. Delivery is complete when every requested assignment is accepted, explicitly deferred, or reported blocked, and all settled owned sessions pass cleanup.