Agent-Skills.md

Agent Skills: Agentic Session Harness (ASH) — engine

Generic, vendor-neutral session-observability engine (ASH — Agentic Session Harness). Per-session journals capturing goal · tasks · decisions[] · sources · transcript_hash, plus a decision-audit (why the agent decided X, spec_alignment drift). Ships CLIs (agentic-walkthrough timeline · agentic-decisions audit report · agentic-decide capture · agentic-reindex backfill) + SessionStart/Stop hooks. Use when you need an auditable record of what an agent did and WHY across sessions. Promoted from a host product 2026-06-02 (Layer-1 community engine).

UncategorizedID: ekson73/multi-agent-os/agentic-session-harness

Repository

ekson73/multi-agent-os
ekson73License: MIT

Install this agent skill to your local

pnpm dlx add-skill https://github.com/ekson73/multi-agent-os/tree/HEAD/skills/agentic-session-harness

Skill Files

Browse the full folder contents for agentic-session-harness.

Download Skill

Loading file tree…

skills/agentic-session-harness/SKILL.md

Skill Metadata

Name
agentic-session-harness
Description
Generic, vendor-neutral session-observability engine (ASH — Agentic Session Harness). Per-session journals capturing goal · tasks · decisions[] · sources · transcript_hash, plus a decision-audit (why the agent decided X, spec_alignment drift). Ships CLIs (agentic-walkthrough timeline · agentic-decisions audit report · agentic-decide capture · agentic-reindex backfill) + SessionStart/Stop hooks. Use when you need an auditable record of what an agent did and WHY across sessions. Promoted from a host product 2026-06-02 (Layer-1 community engine).

Agentic Session Harness (ASH) — engine

Purpose

A session-end observability harness: one journal entry per session (.claude/audit/<YYYY-MM>/<DD>.jsonl) with the 17-field FROZEN schema (goal · tasks · decisions[] · sources · transcript_hash · compliance_flags · …; spec_alignment is an additive field inside decisions[], not a top-level field). The walkthrough reads the timeline; the decision-audit answers why the agent decided X instead of following the SPEC. Vendor-neutral (AAIF) — Bash 3.2 + jq only, no org-specific content (Layer-Purity clean).

When to Use

  • Audit what happened in a session (timeline) → agentic-walkthrough <sid>
  • Audit the agent's DECISIONS + why → agentic-decisions (--filter spec_alignment=divergent = the drift signal)
  • Capture a high-stakes decision mid-session → agentic-decide --decision … --rationale … --spec-ref …
  • Backfill/migrate old journals → agentic-reindex

Trigger Phrases

"session timeline", "why did the agent diverge from the spec", "decision audit", "capture this decision", "walkthrough of session X", "session journal".

Components

| Kind | Path | Role | |---|---|---| | CLI | bin/agentic-walkthrough | session timeline (markdown; --raw, --today, --violations) | | CLI | bin/agentic-decisions | decision-audit report (table/list/json; --filter/--sort) | | CLI | bin/agentic-decide | in-session explicit capture (DEC- ids) | | CLI | bin/agentic-reindex | backfill/migrate/enrich journals | | CLI | bin/agentic-fix-dangling-symlinks.sh | repair transcript symlinks | | hook | hooks/link.sh, hooks/resume.sh | SessionStart: symlink transcript + inject last-5 context | | hook | hooks/stop-fallback.sh | Stop: reliably-firing entry writer + structural XDEC- extraction | | hook | hooks/decide-merge.sh | Stop: fold staged DEC-/XDEC- into decisions[], deduped | | lib | hooks/lib.sh | shared helpers (sourced by hooks + reindex) | | spec | SPEC.md | the FROZEN-17 generic schema SSOT |

The CLIs source hooks/lib.sh relatively (../hooks/lib.sh); the bundle is self-contained.

Activation (opt-in) — wire the hooks

ASH ships as an opt-in capability (not auto-wired, so it never forces journaling on consumers). To activate, register the hooks in your project's .claude/settings.json (SessionStart → link.sh, resume.sh; Stop → stop-fallback.sh then decide-merge.sh), or via the plugin's hooks/hooks.json referencing ${CLAUDE_PLUGIN_ROOT}/skills/agentic-session-harness/hooks/*.sh. Orientation + per-intent routing: see the walkthrough-concierge skill.

Sentinel coexistence (multi-agent-os)

ASH and the Sentinel Protocol both observe into .claude/audit/, but do not collide and are complementary:

  • ASH journals → .claude/audit/<YYYY-MM>/<DD>.jsonl (month-dir + day-file): per-session goal/decisions/timeline + decision-audit (the what & why).
  • Sentinel traces → .claude/audit/session_{YYYYMMDD}_{hex}.jsonl + session_index.json (flat): anomaly/health detection (the is-this-going-wrong). Distinct filename namespaces → both run in the same sink. ASH = post-hoc decision audit; Sentinel = live anomaly guard.

Companions

  • walkthrough-concierge — teaches/routes/anchors over this engine (start here if unsure which tool).
  • decision-capture — WHEN to call agentic-decide.

Changelog

  • 2026-06-02 — v1.0.0 — Promoted to multi-agent-os (Layer-1 community engine) from a host product. Renamed CLIs ash-* → agentic-*; hooks dropped the ash- prefix (dir name covers); self-contained bundle (bin/ + hooks/ co-located, relative lib sourcing). Vendor-neutral / Layer-Purity clean. Corporate overlays (taxonomy/compliance) live separately in the consuming org's private toolkit.