Agent-Skills.md

Agent Skills: Session-Reentry (soul-name **Anamnesis**)

|

UncategorizedID: ekson73/multi-agent-os/session-reentry

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/session-reentry

Skill Files

Browse the full folder contents for session-reentry.

Download Skill

Loading file tree…

skills/session-reentry/SKILL.md

Skill Metadata

Name
session-reentry
Description
|

Session-Reentry (soul-name Anamnesis)

The RECEIVING side of a session boundary. Where postflight emits a handoff-seed at end-of-work (the SENDING side), session-reentry reconstructs and re-onboards at start-of-re-entry. Anamnesis (Plato ἀνάμνησις — the un-forgetting) is the dual of the amnesia every fresh agent wakes into (ai-as-pwd-axiom §1). Design spec: ADR-009 (docs/adrs/).

Naming: the soul-name Anamnesis is operator-overridable per [C-naming] (naming-authority — Anima decides, operator may rename post-hoc). The canonical system-slug session-reentry is what code/routing//maos: key off and does not change on a soul-name override.

Increment status — M1 (text-only walking-skeleton) + M2 (id→path resolution via bin/resolve-session.sh · any-mind register --mind→opera-debrief --audience · dormancy-scaled depth) shipped. M3 audio · M4 graphic + rule cross-refs remain deferred to issue #234 (build-spec: ADR-009 §M2–M4 Build-Spec).

Purpose & the gap it fills

Every session is a fresh amnesic agent. A mind — the original operator, a different human, or another agent — routinely opens an OLD/DORMANT thread after dozens of intervening tasks and must re-attune: rebuild the thread's motivation → purpose → objectives → deliverables → state, then re-enter "the vibe" and continue. That reconstruction is the expensive re-entry cost (Mark 2008: ~23 min to fully resume interrupted work).

A 2-agent prior-art census confirmed the gap: adjacent tools cover the SENDING side (postflight / continuation-seed emit a seed) or the wrong axis — morning-briefing --mode=recap = self + live + sectioned dump · work-compass = the WHAT not the UX · session-report = usage-analytics · ASH = decision-audit · CPT §9 Compass = a cold-rehydration mechanism (no pedagogy). None does cold/foreign RECEIVING + any-mind re-onboarding. Full contrast table: ADR-009 §Context.

This tool = the RECEIVING side: reconstruct a cold/dormant/foreign thread's intent-hierarchy from persisted artifacts, then re-onboard any mind progressively, multi-modally — by composing those primitives, not reinventing them (Strata / Gordian: no new mechanism where one exists; the only net-new are two thin deterministic helpers — the dormancy score + the id→path resolver).

When to use / not use

  • Use: re-entering a dormant/old session; onboarding a different mind (human or agent) to an existing thread; "what was this about, and what's next?".
  • Not use: the session is live and you just need a state-recap → morning-briefing --mode=recap; you're ending work and want to leave a seed → postflight; a single-file Q&A → answer directly.

Flags

| Flag | Default | Meaning | |---|---|---| | --session <id\|path> | current | target thread; an id resolves via bin/resolve-session.sh (M2) | | --mind self\|human\|agent | self | audience register → opera-debrief --audience (M2) | | --depth auto\|L1\|L2\|L3\|full | auto | progressive layer; auto = the dormancy band; expand-on-demand | | --media text\|audio-voice\|graphic | text | delivery modality (audio/graphic opt-in; M3/M4) | | --scope current\|down\|sideways\|up\|forward | current | CPT Compass verb (§9) |

M2 wiring (this increment) — --session <id> resolves to an artifact path via bin/resolve-session.sh (an existing path passes through; no match → honest non-zero, never a fabricated path). --mind maps to the RE-ATTUNE register: self/human → opera-debrief --audience human (warm; self at lower --intensity), agent → --audience agent (machine-economy digest). --depth (defaulting to the dormancy band) gates how many N-Tree layers render (L1→L2→L3→full). --media audio/graphic remain M3/M4 — they emit a stderr [--media] audio/graphic deferred to M3/M4 (#234) + fall back to text (no fabricated modality, anti-theater R3).

Architecture — 3 phases, each COMPOSING primitives

Scaled by the dormancy score (bin/dormancy-score.sh — a net-new bit): staler/more-foreign ⇒ deeper reconstruction (grounds Monk-Trafton resumption-lag).

| Phase | Does | COMPOSES (reuse — do not rebuild) | |---|---|---| | 1 · INGEST | resolve --session id→path (bin/resolve-session.sh); walk persisted surfaces; compute dormancy | CPT Compass walk (akasha rules/cowork-process-topology-protocol.md §9: INDEX.yaml/manifest/events) · ASH (skills/agentic-session-harness) · session-report/analyze-sessions.mjs (transcript parse) · skills/postflight/references/continuation-seed-contract.md (seed shape) · ticket-as-prompt re-entry-block (akasha) · the 37-type work-surface taxonomy (akasha docs/work-surface-taxonomy-2026-07.md) as the input-surface map | | 2 · RECONSTRUCT | build the objectives N-Tree (originary/primary/secondary/auxiliary) + deliverables + state | skills/postflight P2 DEBRIEF (N-Tree synthesis — the owner) · morning-briefing --mode=recap (state-recap) · skills/work-compass (cross-surface aggregation) · skills/pulse (re-orientation phasing) | | 3 · RE-ATTUNE | layered, any-mind onboarding (--mind→opera-debrief --audience; --depth→layers) | skills/opera-debrief (human/agent register + audio passthrough) · skills/content-recast (progressive-disclosure lens + faithfulness guard) · producers (M3/M4): skills/voice-director+bin/speak.sh (audio), content-recast render → NotebookLM/Gamma (graphic), session-report/template.html (HTML) |

Dormancy score + id→path (the net-new bits)

bin/dormancy-score.sh <session-path> → a 0.00–1.00 scalar + JSON envelope; the sibling bin/resolve-session.sh <id|path> turns a session id into that path (M2 — closes the M1 "the bin takes a path only" gap; an existing path passes through). Dormancy inputs: newest-artifact mtime staleness + a foreignness signal (is the re-entering mind the author?). Output contract:

{ "score": 0.67, "band": "deep", "age_days": 14, "foreign": false,
  "recommended_depth": "L3", "signals": { "mtime_days": 14, "foreign_author": false } }

Bands (scale reconstruction depth): fresh <0.25 → L1 · warm 0.25–0.55 → L2 · deep 0.55–0.80 → L3 · cold >0.80 → full. Both bins are deterministic (same inputs → same output); no interpretation-dependent output (mirrors the Convergence-Engine deterministic-harness discipline).

Output layering (the pedagogy — research-grounded)

Endsley L1 perception (artifacts/state) → L2 comprehension (the objective N-Tree = "the vibe") → L3 projection (next-actions), nested inside Nielsen progressive disclosure (reveal-on-demand), chunk-gated ≤~4–7 units (Miller / Sweller), opened by an Ausubel advance-organizer (the 🌱 originário root), wrapped in the corpus sign-system (🌱 originary / 🌳 N-Tree), closed by exactly ONE primary recommended next-action (the CTA). Default render = the --depth / dormancy band; the mind expands on demand.

Output contract (single-CTA — disambiguated)

Re-entry renders one primary recommended action as the CTA. The Handoff Option Menu (up/down/sideways/forward Compass verbs) is secondary and optional — shown only on request or at --depth full. Never dump every option as co-equal (that is the "spit everything at once" failure this tool exists to avoid).

Faithfulness guard (anti-theater R4)

Reconstruct ONLY from persisted artifacts. Where an artifact is missing/silent, render an honest "unknown / not persisted" — never a confident guess (garbage-in is surfaced, not fabricated). Low artifact-coverage ⇒ shallower, flagged output. Both bins fail clearly (non-zero) rather than fabricate a path/score.

Definition of Done

  • M1 ✅ — 3-phase reconstruction from persisted artifacts · progressive layered output (L1→L2→L3) · one CTA + optional Compass menu · faithfulness guard · every phase composes an existing primitive except the dormancy score · SKILL + bin + smoke + merged (PR #235).
  • M2 ✅ — id→path resolution (bin/resolve-session.sh) · any-mind register (--mind→opera-debrief --audience) · dormancy-scaled depth · smoke assertions.
  • M3/M4 — deferred to issue #234 (build-spec: ADR-009 §M2–M4 Build-Spec).

Anti-patterns (do NOT)

  1. ❌ Spit-everything — dumping the full transcript/all options at once (defeats progressive disclosure + the single-CTA contract).
  2. ❌ Fabricate the un-persisted — inventing motivation/objectives (or a session path/score) an artifact doesn't support (anti-theater R4).
  3. ❌ Reinvent a primitive — re-implementing N-Tree synthesis (postflight P2 owns it), state-recap (morning-briefing), or the register/audio (opera-debrief owns them) instead of composing them (Strata).
  4. ❌ Claim a modality it can't render — audio/graphic (M3/M4) degrade with an honest stderr note (anti-theater R3), never a faked artifact.
  5. ❌ Interpretation-dependent bins — dormancy + resolver must be deterministic.

Quality tests (self-validity)

  1. Self-Application — re-entry can reconstruct its own build session. ✅ (dogfood)
  2. Non-Contradiction — composes postflight/morning-briefing/work-compass/opera-debrief without duplicating; distinct axis (RECEIVING vs SENDING). ✅
  3. Survival — advocates faithful reconstruction; renders honestly on itself. ✅
  4. Bounded — id→path + register + depth scope · deterministic bins · Goldilocks <12KB · #234 defers M3/M4. ✅
  5. Explicit-Exception — faithfulness guard + degraded-modality note + --depth bounds. ✅
  6. Utility-Sunset — deprecate if a host ships native cold-thread re-onboarding (DUED). ✅

Versioning

  • v0.2.0 (2026-07-14) — M2: bin/resolve-session.sh (id→path resolver — closes the M1 path-only gap; deterministic, honest-on-no-match) · --mind → opera-debrief --audience register wiring · --depth gates N-Tree layers from the dormancy band · smoke assertions for the resolver. M3 (audio) / M4 (graphic + rule cross-refs) stay deferred to #234; build-spec = ADR-009 §M2–M4.
  • v0.1.0 (2026-07-10) — M1 walking-skeleton (PR #235): 3-phase reconstruction + bin/dormancy-score.sh + smoke; text-only, single-CTA, faithfulness guard.

Related

  • docs/adrs/ADR-009-session-reentry-anamnesis.md — the design spec (+ §M2–M4 Build-Spec).
  • skills/postflight — the SENDING dual (P2 DEBRIEF / P3 HANDOFF seed this RECEIVES).
  • skills/morning-briefing · skills/work-compass · skills/pulse · skills/content-recast · skills/opera-debrief · skills/agentic-session-harness — composed primitives.
  • Build tracker: ekson73/multi-agent-os#234.

License

MIT (matches multi-agent-os repo LICENSE).