port-daddy

Instruction manual for agents driving Port Daddy multi-agent coordination. Use when an agent will edit a repo, recover work, coordinate with other sessions, inspect FleetBar/Fleet Control Center truth, package skill/docs surfaces, or leave a durable handoff. NOT for generic coding that does not need Port Daddy state.

Contributor manual for agents working ON the Port Daddy codebase itself — the daemon, MCP server, FleetBar / Fleet Control Center, website, CLI surface, distribution mirrors, internal recovery ledger, and the named internal actors (Coxswain / Navigator / Cartographer / Lookout / Quartermaster + Shipwright). Use when editing the port-daddy repo. NOT for agents using Port Daddy on other projects (use port-daddy-agent-skill for that), and NOT distributed to public skill catalogs — this skill is private to the port-daddy repo.

Decide which single operator surface — Scout, FleetBar, or pd-console — owns each capability by its distance-from-work (intake/ambient/deep), and audit that placement for authority spread, unenforceable controls, evidence overflow into FleetBar, and hot/cool bus-subscription mismatches. Use when placing a new capability on one of Agent Harbor's three operator surfaces, reconciling a mockup that duplicates a capability across two surfaces, or auditing an existing operator-surface spec before implementation locks it in. NOT for choosing SDK/CLI/MCP/GUI surfaces for API-consuming developers (developer-surface-strategist), designing the concrete interaction flow within one already-assigned surface (agentic-coding-ux-designer), or the hot-bus/cool-bus transport mechanics themselves (swarm-invocation-designer).

Audit what PRs this session produced. Ask: "What work did I do this session that isn't in a PR yet, or isn't merged?" Forces the agent to account for all code changes before declaring done. Use at any point — especially at session end, after a manager wave, or when asked "what's left?"

After each execution wave completes, inspect the DAG's commitment landscape and premortem risk score. If any surviving nodes carry `commitment_level: TENTATIVE`, or if the premortem `recommendation` is `ACCEPT_WITH_MONITORING` or `ESCALATE_TO_HUMAN`, pause execution and run a structured parley: re-evaluate TENTATIVE nodes against the evidence produced by the just-completed wave, update risk severity where warranted, and either promote nodes to COMMITTED, demote them to EXPLORATORY, or prune them before launching the next wave. Parley is a scheduled operation triggered by wave completion — not an ad-hoc intervention — making wave boundaries the natural formation-break point where plans meet reality.

Build and extend pd-console — Port Daddy's GPU-native macOS operator console (GPUI 0.2.x, Zed's Rust UI). Covers the render-agnostic Block/Pane(Surface) contract, the two-thread reqwest↔smol refresh pipeline, Taffy flexbox layout, uniform_list virtual scroll, focus + keyboard nav, the OKLCH theme and ICS maritime flag badges, GPUI's missing text-input, and the real feature-gated cargo/CI gate. Use when adding panes, visual polish, or debugging GPUI rendering/layout/focus in core/pd-console. NOT for the TypeScript daemon, generic Rust toolchain/borrow-checker help (use rust-with-claude-code), or non-pd GPUI apps with a different theme/architecture.

Chooses the most important product or roadmap work under ADHD, founder, budget, and multi-agent constraints. Use when asking what to work on next, prioritizing a roadmap, cutting scope, choosing an MVP slice, or deciding whether to split work across agents. NOT for executing an already-chosen task, generic project status reporting, personal financial advice, or long unbounded brainstorming.

Algorithmic skill for optimally partitioning LLM context or task DAGs across K agents. Covers two modes: (A) online/reactive spawning under context pressure, and (B) static pre-planned DAG execution with graph min-cut optimization. Encodes algorithms for sequence-stable clustering, dynamic K selection, token-budget-aware partitioning, and handoff protocol design. Use when designing multi-agent systems, deciding when and how many agents to spawn, or optimizing a WinDAGs-style DAG for token efficiency.

Design discovery, directories, and guilds for agent swarms at scale. Use when many agents (5 -> hundreds) must FIND each other on a shared surface: capability indices, yellow-pages directories, whois/phonebook routing, reputation-aware discovery, and guild structures that scope trust. Covers the canonical prior art (FIPA Directory Facilitator, Contract Net, Chord/DHT, DNS-SD/Consul, EigenTrust, Greif's merchant guilds) and the failure modes (read-poverty, Sybil reset, Goodhart on rank, directory staleness, cold-start). Triggers: "agents can't find each other", "who owns this file/skill", "route a question to the right agent", "capability registry", "agent reputation", "guild / federation of fleets".

Install Port Daddy coordination hooks ("tentacles", from the Giant Squid Harness) into the INTERACTIVE sessions of agentic coding CLIs (Claude Code, Codex, Gemini, Antigravity/agy) by auto-detecting which CLIs are installed and writing PER-PROJECT hook config behind a runtime gate, so coordination fires only inside a Port Daddy project while the daemon is running — never machine- wide-always-on. Use when the user says "install port daddy hooks", "agent CLI hooks", "interactive session hooks", "make coordination hooks fire in interactive sessions", "codex/gemini/agy hooks", or "auto-detect agent CLIs". NOT for generic software/package installation; NOT for installing the Port Daddy daemon or `pd` CLI itself (delegate to the `port-daddy` skill / `pd setup`); NOT for MCP server setup (delegate to `pd mcp install`); NOT for authoring new skills (delegate to `skill-architect` / `skill-creator`); NOT for editing the tentacle scripts themselves (owned by the Giant Squid Harness program; this installer layer is pd-ad

Conduct Critical Decision Method (CDM) interviews to elicit expert reasoning strategies (L3 knowledge) from practitioners about specific challenging past incidents. Use when capturing expertise for skill development, preserving knowledge before practitioner departure, building post-incident learning artifacts, or investigating decisions in complex sociotechnical systems. Outputs structured incident reconstructions with cues, decisions, alternatives considered, and counterfactual analysis suitable for distillation into skills or training material.

Render mathematically-faithful Turkish marbling (Ebru) and paper-marbling with closed-form coordinate-deformation maps — a drop is a radial outward push, a comb/tine is a decaying shear, all composed as homeomorphisms over colored region boundaries. No fluid solver, runs 60fps on a 2D canvas. Use when building a marbling/Ebru toy or generative-art renderer, implementing drop/comb/wave/swirl deformations, scripting traditional patterns (Battal, Gel-Git, Şal, Bülbül Yuvası), keeping combed edges crisp via adaptive resampling, modeling ox-gall shouldering for crisp ring-in-ring drops, or exporting high-DPI marbled prints. NOT for real fluid / Navier-Stokes simulation, metaball / alpha-blended "smoke" blob effects, raster image-processing filters, or non-marbling generative art.

Treat tokens as a coding swarm's cost-of-goods-sold AND its legibility engine. Covers per-agent context budgeting, compaction strategies (in-context summarization, structured note-taking, sub-agent isolation), the shared digest as compaction-for-humans- and-agents, the context-degradation cascade (lost-in-the-middle, context rot, recursive- summarization collapse), and mechanism design for metering token spend across a fleet. Activate on: "token budget", "context budget per agent", "compaction strategy", "digest / briefing as compression", "context window degradation / context rot", "summarization collapse", "COGS for agents", "metering agent token spend", "pheromone decay as forgetting", "/context-economics-for-agent-swarms". NOT for: memory storage architecture (use always-on-agent-architecture), single-prompt wording (use prompt-engineer), pure mechanism-design proofs (use nisan-et-al-2007-algorithmic-game-theory).

Produce academic LaTeX whitepapers that look expensive, not cheap. Use when writing or revising a technical whitepaper (.tex), designing its TikZ figures, or reviewing one that "looks cheap." Encodes a restraint-first house style (one serif typeface, one accent color, consistent figure craft), the anti-cheap checklist, the pedagogic apparatus (callouts, exercises, reader's-map, pull-quotes), and a mandatory render-read-audit loop.

Make a swarm of autonomous coding agents LEGIBLE to one human operator without high-modernist over-flattening. Applies James C. Scott's *Seeing Like a State* (legibility, mētis, high-modernism) and Hobbes' *Leviathan* (state of nature -> consented authority) to multi-agent software systems. Core discipline: DIGEST-WITH-ZOOM — every summary is a lens onto the real artifact (diff, reasoning trace, claim row), never a replacement for it. Use when designing read-surfaces, dashboards, attention queues, operator consoles, briefings, resurrection digests, or any "what is the swarm doing?" view; when an operator is drowning in illegible agent output; when deciding what to summarize vs. preserve; or when an oversight surface risks putting the human OUT of the loop. NOT for: choosing coordination topology (use multi-agent-coordination / Conway), the wire protocol of agent messages (use agent-conversation-protocols), or the economics of agent labor (use mechanism-design-for-agent-labor).

End-to-end LoRA/QLoRA fine-tuning for open-weight models in 2026 — picks a base model from its strengths/weaknesses, assesses whether to train locally or in the cloud and sets up either, runs the training, and visualizes both the dataset and base-vs-tuned outputs. Use when fine-tuning, training a LoRA/QLoRA/DoRA adapter, choosing a base model, deciding local-vs-cloud GPU, or previewing training data and results. NOT for full-parameter pretraining, closed-model API fine-tuning (OpenAI/Gemini), dataset curation from scratch (use fine-tuning-dataset-curator), or RAG/prompt-only work.

Implement macaroons (Birgisson et al. 2014) as unforgeable, attenuate-only bearer capability credentials, with cross-language byte-parity. Use for capability tokens, bearer-credential attenuation, third-party caveats + discharge gates, delegation chains, "rent-paid"/compulsion caveats, or keeping a Rust and a TypeScript macaroon impl byte-identical via shared test vectors. Covers HMAC-chained first-party caveats, the two vid (verification-id) constructions (HMAC commitment vs AES-GCM sealing) and when to use each, per-hop verification soundness, request-binding, constant-time + fail-closed verification, and the structured caveat grammar. Pairs with agentic-zero-trust-security and rust-kernel-ffi. NOT for: OAuth/OIDC/JWT web-session auth, centralized token servers, or confining a malicious same-UID process (that needs OS/VM isolation — macaroons make the gate unforgeable, not the holder confined).

Construct PowerPoint, Keynote, Reveal.js, or Marp decks that are both legible from the back of the room AND feel premium / crafted / expensive. Use for pitch decks, interview decks, board decks, conference talks, slidedocs, investor memos, lecture slides. NOT for logo / wordmark design (→ brand-guidelines), copywriting (→ copywriting), motion graphics for video (→ motion-design-web), live-stage stagecraft (out of scope), or live-presentation coaching (→ tech-presentation-interview).

Design agents that are solely responsible for one unique concern of a project — project avatars, watch officers, wardens, sentinels — with exclusive scope, a durable state home, a mandatory ledger, an escalation path, and a handover protocol. Use when creating a per-project guardian or log-watching agent, when promoting an existing fleet agent (test-hunter, spark, spider) to true sole ownership of its concern, or when choosing where a long-lived agent's state should live (notes vs tuples vs episodes vs own tables). NOT for one-shot task agents, general multi-agent orchestration topology, skill authoring, prompt engineering, or CI pipeline design.

Design THREE-sided agent-labor markets: operators selling labor + fleet for-hire, agents/fleets as rentable assets, skills/tools licensed — all on one bond/credit ledger with float-plan escrow. Use when the market has more than buyer+worker (a platform/operator side, an asset-rental side, and a skill-licensing side), when designing "hosted trust" pricing (you sell the verified ledger + relay + reputation, not the crypto), or when reasoning about cross-side network effects, price-structure (who subsidizes whom), and identity→continuity→reputation→tradeable-asset chains. NOT for two-sided requester/worker bond pricing in isolation (use mechanism-design-for-agent-labor), not for generic DeFi tokenomics, not for payment-gateway plumbing.

Runs CDM (Critical Decision Method) interviews to elicit the perceptual layer missing from science-only SKILL.md files. Activate when: Looking Back Q3 fires with scope≠narrow, a pattern_crystallization event is detected, a skill underperforms (expectancy_violation), or a novel task has no catalog match. Produces structured SKILL.md v2 metadata: recognition-cues, expectancies, decision-cues, adaptive-workarounds. NOT for: general Q&A, skill selection, DAG execution.

Design and scaffold new skills, subagents, and MCP surfaces. Use when choosing which primitive to create. NOT for invoking existing agents, generic coding help, or deployment work.

Design scheduling agents. Use for timezone-safe negotiation, focus protection, and privacy-safe free-busy handling. NOT for calendar UI, project planning, or general agent architecture.

Apply AlphaGo-style patterns: learned priors plus search, mixed evaluators, self-play, and asynchronous fast/slow loops. Use for intractable search and compute allocation. NOT for basic supervised learning, tractable optimization, or domains without search bottlenecks.

Analyze repeated non-engagement, deflection, motte-and-bailey shifts, tone policing, sealioning, and related manipulative discourse patterns. Use when the user suspects they are being rhetorically managed rather than answered, or when they need a careful pattern read grounded in quotes and alternatives. NOT for psychiatric diagnosis, legal judgment, or help deploying bad-faith tactics.

Russell Beale's "Rise up, Revolt!" analysis of why users blame themselves for bad design, and how consciousness-raising (not designer education) is the lever that fixes usability at a market level

Representations for hierarchical skill structures including knowledge graphs and ontological decomposition

Use for affordances, hierarchical control, nullspace composition, and information-seeking action selection in embodied agents. NOT for pure symbolic planning, generic RL tuning, or domains without sensorimotor grounding.

Use when planning, scheduling, state variables, temporal constraints, bottlenecks, or commitment-level decisions are entangled in one constrained system. NOT for lightweight to-do planning, generic project-management advice, or domains that lack explicit causal and resource constraints.

Cross-cultural communication frameworks for understanding discourse patterns across cultures

Comprehensive benchmark suite for evaluating LLM agents across diverse interactive environments

Identify, label, and explain logical fallacies in arguments and discourse. Use when the user wants to analyze arguments, check reasoning, evaluate debate claims, spot weak arguments, do debate prep, policy writing, or conflict analysis. Triggers on 'fallacy', 'logical error', 'faulty reasoning', 'argument flaw', 'is this valid', 'check this logic'. Works on both formal deductive errors and informal reasoning mistakes.

SHOP2 hierarchical task network planner for automated planning through recursive task decomposition

Facilitate collaborative, emotionally aware conversations and guide toward productive dialogue. Use when the user wants de-escalation, mutual understanding, structured discussion, conflict resolution, or mentions 'productive discussion', 'collaborative debate', 'dialog vs debate', 'find common ground', 'help us talk about', 'de-escalate'. Includes a couples-in-recovery mode for fragile nervous systems. Can orchestrate other discourse skills in a structured sequence.

Guide critical thinking through disciplined Socratic questioning. Use when the user wants to explore assumptions, examine beliefs, resolve internal conflict, facilitate learning, clarify thinking, or mentions 'Socratic method', 'question assumptions', 'examine this belief', 'think this through', 'interview my beliefs'. Especially powerful for relationship processing, recovery decisions, and examining deeply held assumptions.

Construct the strongest, fairest version of someone's position before engaging with it. Use when the user wants to understand opposing viewpoints, prepare for debate, ensure fair representation, or mentions 'steel man', 'strongest argument', 'best case for', 'charitable interpretation'. Also useful for relationship conflicts where understanding the other side genuinely matters.

Structure and analyze arguments using the Toulmin model (Claim-Data-Warrant-Backing-Qualifier-Rebuttal). Use when the user wants to build logical arguments, evaluate reasoning, prepare formal debate, clarify messy reasoning, write persuasive pieces, or mentions 'argument structure', 'warrant', 'backing', 'logical framework', 'structure this argument'. Useful as input for speech-writing, fallacy detection, and steelmanning.

Apply AgentSpeak(L) BDI architecture to event-driven autonomous agents with persistent goals, context-sensitive plans, and explicit selection functions. Use when designing interrupts, scheduling policies, or multi-agent coordination without a central controller. NOT for purely reactive rule systems, static planners, or ordinary ETL-style pipelines.

Foundational concurrent computation model where actors communicate exclusively through asynchronous message passing

BDI (Beliefs-Desires-Intentions) agency framework for designing autonomous agents with mental state architectures