aigent-os

The capsule verb, invoked explicitly — reconcile from the live memory, write a resume-ready capsule, then stop. Mid-session checkpoint or completion capsule; a rolling best-effort version of the same write also runs automatically on every Stop event via daemons/stop-capsule-writer.mjs.

The resume verb, invoked explicitly — load the newest valid capsule by created_at, re-ground against live memory, then ACT on waiting_on. Mid-session or on explicit ask; the post-clear boot fires the same procedure automatically via daemons/resume-verb.mjs.

Run the volunteering ledger before declaring work done. Forces surfacing of what changed/didn't, what was noticed-but-not-fixed, what was guessed, and what tradeoffs were made on the principal's behalf. The standard - the principal should never be surprised later by something the agent knew at the time.

Capture a confident agent claim (Phase 1) or resolve one as right/wrong (Phase 2) in vault/memory/TRUST_DECAY.md. Builds the longitudinal record of agent calibration over time. Pairs with /honesty-check (which should reduce Phase 1 frequency) and /diagnose (which surfaces causes that often reveal Phase 2 events).

One evidence-gated pass that sequences /dream, /reconcile, /sweep-now, and /digest over your own vault, with a /system-check gate and a ledger-review pass over the measurement layer (Honesty Ledger, Trust Decay, Failure Modes). Stays exactly as human-gated as the legs it calls -- stages, never auto-applies. Skips a leg gracefully when there is nothing to do; every leg that runs leaves an evidence line.

Dispatch Hestia to run all 3 vault sweeps (DELEGATION_TRACKER stale items, HEAT_INDEX dormant flips, broken wikilinks) and append the result to HESTIA_SWEEP_LOG.

Route ONE bounded, mechanical task to the Codex CLI as a non-Claude executor, under the same review-before-merge gate as any other agent-produced diff. Use when the user explicitly asks to hand a task to Codex, or names OpenAI Codex / a non-Claude executor for a mechanical change (rename a symbol across files, add a config entry, apply a scripted find/replace). Not for open-ended or strategic work -- that stays with the operating Claude session.

Mute a Caddy hint class for a window. Defaults to 4h auto-expiry. Classes — memory, context, body, routing, all.

The day-one first-run flow and the everyday home screen. On first launch, greet, run setup, then carry the operator into a real first win. On return, give a short briefing and action menu.

Commit the session to durable vault memory and prepare a clean resume point

Boot the session, load context from the vault, and surface what matters

Configure or reconfigure aigent-OS identity, priorities, authority, projects, people, and decision logic

Track and surface sub-agent dispatch calibration ratios. Extract from JSONL transcript or read existing ledger. Per-agent clean/blocked/errored/partial trends.

Compose the AIgent's body state by reading existing signals. Lazy compute. Never polled. Reports the five pressures + recommended reflex.

Generate a structured delegation brief for any task

Quarterly skill that walks the skill-index against the actual SKILL.md files on disk, reports drift (skills indexed that no longer exist, skills on disk that aren't enrolled, skills whose descriptions have changed since enrollment). Hardens the auto-enrollment pipeline against silent rot.

Enroll a skill into Caddy's index so it can be surfaced when the right prompt arrives. Use when a new skill is created, copied in, or when a skill needs updated triggers. Usage — "/caddy-enroll <path to SKILL.md>".

When a [CADDY] hint fires and the principal is confused why, /caddy-explain walks the scoring transparently — which trigger matched, what points it earned, what the threshold was, what other skills were close. Pairs with the legibility thesis. Caddy's decisions should be inspectable on demand.

Compact a long capsule chain. Walks parent_capsule_id backward from a head capsule; if chain >= 5, summarizes oldest 3 into a chain-summary capsule. Manual invocation.

Echo's comms polling skill. Read an inbox, channel, or message thread and return a structured unread summary — sender, topic, action required, and urgency. Read-only.

Mnemosyne's note compaction skill. Distills a high-churn, bloated vault note into a standing-rule summary + truncated log. Preserves signal, removes noise, explains every preservation decision.

Hypatia's plan critique skill. Takes a strategic plan, architectural decision, or proposal and returns adversarial pre-commitment review — strongest counterargument first, hidden assumptions surfaced, alternatives named.

Run a decision through the 12-lens framework

Run extended multi-agent reconnaissance sessions. Use when asked to brainstorm deeply, explore ideas from multiple angles, or generate a structured recon document.

When something feels off, type /diagnose and describe the symptom. The skill walks the seven-layer stack (kernel, skills, hooks, daemons, vault, doctrine, principal config) and reports which layer is most likely responsible. Caddy pattern applied to debugging.

Review staged memory candidates from MEMORY_CANDIDATES.md. Surface each with promote/skip/supersede options. Does not auto-promote.

Classify any incoming directive as Ship / Ask one focused question / Propose first. Forces a sharp decision instead of a multi-option survey to the principal. The default failure mode for agents is treating ship-work as ask-work (over-surveying) or treating propose-work as ship-work (rushing irreversible changes). This skill prevents both.

Hypatia's targeted weakness-finding skill. Given a specific claim, architecture, or decision, finds the single strongest objection and the hidden assumption most likely to be wrong. Sharper and faster than critique-plan.

The exact operational recon sequence for an unfamiliar codebase or unfamiliar area. Replaces "ask the human to clarify" with "read the code first." Most clarification questions are answered by 5 minutes in the relevant file - this skill enforces reading before asking.

The day-one first-win menu. Three dead-simple, high-value tasks (two zero-dependency) that each put a real, usable artifact in the operator's hands in minutes, so the "ship something real" promise is met on day one and never waits on a connector or OAuth.

Create stunning, animation-rich HTML presentations from scratch or by converting PowerPoint files. Use when the user wants to build a presentation, convert a PPT/PPTX to web, or create slides for a talk/pitch. Helps non-designers discover their aesthetic through visual exploration rather than abstract choices.

Transform AI-generated documentation into human-readable prose with personality. USE WHEN: Converting rigid AI docs (checklist spam, CAPS LOCK emphasis, perfect symmetric structures, template dumps) into flowing, conversational writing that feels colleague-authored. Different from /humanizer — that skill detects AI writing patterns (Wikipedia taxonomy). This skill does aggressive deformalization of document structure and flow. PRIMARY TRIGGERS: "humanize this doc", "make this readable", "remove AI patterns from doc", "add human touch to this" See also: humanizer skill for text-level AI pattern removal.

Create distinctive, production-grade frontend interfaces with high design quality. Generates creative, polished code that avoids generic AI aesthetics. Use when the user asks to build web components, pages, artifacts, posters, or applications, or when any design skill requires project context. Call with 'craft' for shape-then-build, 'teach' for design context setup, or 'extract' to pull reusable components and tokens into the design system.

When the principal is in learning mode (reading a repo, watching a talk, processing a new tool), /learn takes a brief input and produces a structured note in vault/concepts/learning/ with category, modern-stack location, relevance verdict, and reconsideration triggers. Captures the value of every learning sweep without requiring it to immediately translate to building.

The AIgent first-run operator-business interview. Learn the operator's business in three plain questions, write the first memory, and play back a briefing that reads personal, not templated. This is the "it remembers me" moment. A distinct skill that LAYERS ON TOP of the base /setup (which is left intact), it does not replace it. The first-run flow (/start) calls this, not the generic /setup.

Lyra's spec-port skill. Takes a complete written specification and ports it into working code or a vault artifact. Confirms scope, reads the target, builds, returns honesty ledger.

Take a slice of the current conversation and promote it to a durable vault artifact — concept note, project update, person profile, or doctrine entry. Pairs with /close. /close writes the daily note (transient outcome record); /promote elevates specific insights to permanent vault content the framework can refer back to.

Newton's multi-source deep research skill. Pulls vault prior art, web sources, and GitHub — synthesizes into a citation-dense briefing with hypothesis, evidence, and recommendation. Returns honesty ledger.

SELF-CORRECT tool choice. About to Read a long file just to extract one piece of info? Use a sandbox tool that returns only the summary. About to run a long shell command? Same. Sandbox keeps raw output OUT of context — only summaries land. Triggers when about to flood context with large output.

Echo's vault traversal skill. Read-only graph walk — find all notes related to a topic, trace wikilink chains, inventory a directory, or spelunk a codebase for definitions and call chains.