[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M757W981Z9E3FBCW4MA5PEFD":3,"skill-same:01M757W981Z9E3FBCW4MA5PEFD":79},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":16,"categories":17,"keywords":18,"publishedAt":19,"linkOnly":20,"markdownContent":21,"markdownBody":22,"frontMatter":23,"refName":30,"commitSha":31,"skillDirectory":32,"forksCount":33,"repositoryHtmlUrl":34,"repositoryHomepageUrl":35,"ownerHtmlUrl":36,"ownerBlogUrl":6,"ownerDescription":37,"updatedAt":38,"install":39,"resources":52},"01M757W981Z9E3FBCW4MA5PEFD","agent-harness-design",null,"Use when designing a coding-agent harness or runtime.","lukemcqueen\u002Fsteadfaste-fleet","lukemcqueen","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F405240?v=4","01M757MHPAGRXTJAT46SKZP98R","Luke McQueen","01M757MHPZ573KP8DT9GB8SBQ7",4,"MIT","https:\u002F\u002Fgithub.com\u002Flukemcqueen\u002Fsteadfaste-fleet\u002Fblob\u002Fmain\u002Fskills\u002Fdevops\u002Fagent-harness-design\u002FSKILL.md",[],[],"2026-11-12 13:55:43",false,"---\nname: agent-harness-design\ndescription: \"Use when designing a coding-agent harness or runtime.\"\nversion: 1.0.0\ncategory: devops\nauthor: Steadfaste Fleet\nlicense: MIT\nplatforms: [linux, macos]\n---\n\n# Agent Harness Design — Boring, Stable, Agentic\n\nDesigning a coding-agent harness, runtime, or orchestration layer. Luke's #1\nrequirement is **boring stability**: set-and-forget, pinned versions, no\nauto-updates, rare deliberate updates. A foundation that sits unchanged for a\nyear and keeps working.\n\n## When to Use\n\n- Building or evaluating a coding-agent harness, runtime, or multi-worker orchestration layer\n- Deciding whether to adopt vs. fork vs. reimplement an upstream agent framework (Pi, OpenSwarm, OpenHands, Claude Code, Manus)\n- Adding a worker, tool gateway, checkpointing, task queue, or TUI cockpit to an existing harness\n\n## Core principle: the agent is GOVERNANCE around the model (risk + cost)\n\nDo not design a harness as \"core + a worker that does the work.\" The agent is\n**intelligent governance wrapped around a model\u002Fbrain** — and the governance is\nthe product, not the model. The layer between an inbound request (a chat\nmessage) and the executor decides, per request: budget + wall-clock deadline,\nwhich model (cheapest-first escalation), which tools it may touch, what needs a\nhuman's approval, which template\u002Frecipe to follow, when to retry\u002Fescalate\u002Fstop,\nand records every decision. That is **risk management** (what it may\u002Fmust-not\ndo, approvals, audit) + **cost management** (model choice, budget, cheap-first).\nEnterprise adoption is, in the end, just *less risk* — and the yield of the\nwhole product is *repeatable quality*: a variable-quality model producing the\nsame good result, run after run, then improving via a self-healing\u002F\nself-learning loop that pulls today's lesson into tomorrow's policy.\n\nDesign implication: model the **governor\u002Fpolicy engine** (business-editable\nrules, templates, cost\u002Fmethod intelligence) as a first-class user\u002Fuser-editable\nlayer, and hand the executer a *policy* (not just an objective) — the solid\nTask carries budget, permissions, model policy, verification policy, and a\nwall-clock deadline. This governance layer is what serious harnesses famously\nlack; do not forget it and build just a worker dispatch.\n\n## Structure the repo first, and OVER-structure it (before any real code)\n\nRestructure\u002Frename\u002Fmove the repo layout **early**, while nothing real depends on\nthe paths — the user's explicit call is to over-structure with the future in\nmind. A folder or file name that will not survive contact with real code is a\nsign the design is not done. When the settled architecture differs from the\nscaffold on disk (e.g. a TypeScript prototype sitting in a `core\u002F` that is about\nto hold Rust), rename\u002Fmove it immediately so browsing the tree is not\nmisleading — never leave legacy-code-as-current mislabeling. Lay down stubs and\ninterfaces first; a well-named public interface *is* the architecture, and it is\nwhat a future session (or agent) reads cold. File names and code names must be\nintelligent for human review — a name that reads like an index beats a name that\nneeds the file opened.\n\n## Settle the solid contract's decisions BEFORE writing it\n\nDo not write the ABI\u002Fcontract code (types, structs, enums) until the contract's\n*open decisions* are pinned. For a long-term-stability ABI (3 years, rarely\ntouched) the cheap-now\u002Fexpensive-after-it-is-solid fixes matter: the permission\nvocabulary (enumerate the strings and define default-deny for unknowns), the\nwall-clock deadline field, and additive-only growth (new variants ship\nalongside, never replacing). Draft the decision doc + get sign-off, THEN encode\nit in the target language. Writing the solid contract ad hoc hardens undeclared\nchoices outright.\n\n## Core principle: core defines reality, workers adapt\n\nThe runtime's vocabulary is its own and never changes. It knows `task_id`,\n`attempt_id`, `worker_id`, `workspace_id`, `checkpoint_id`, `lease_id` — and\nNEVER learns what a \"Pi session\" or \"Claude conversation\" is. Workers, models,\nand agent frameworks are disposable adapters plugged in behind a solid\ncontract. A three-year-old worker must still connect.\n\n## Core principle: the governance IS the trust layer — give it covenant vocabulary\n\nThe governance\u002Fpolicy engine is not plumbing; it is the product, and it earns\nits own **covenant-toned vocabulary** that a non-technical business owner can\nread at a glance. Luke iterates on this vocabulary directly and cares deeply\nabout the words — expect him to veto and rename terms, and drive each rename to\nits settled form rather than defending a proposal. Established (settled) set:\n**Constitution** = the solid core \u002F supreme law everything runs inside;\n**Covenant** = the bond of trust (the core is that trust *written as law*);\n**Charter** = the rules you set; **Mandate** = one job's orders;\n**Steward** = the human (cares for + is answerable for the system — chose over\n\"Guardian\" because a guardian watches, a steward owns the call);\n**Agent** \u002F **Worker** = governed actor \u002F clinical motor, both swappable;\n**Ledger** = the unchangeable record. Keep universal industry words (Agent,\nTask, Budget, Audit, Retry, Timeout) as-is for coherence; only novel\ncomponents get covenant-toned names. Register preference: the user dislikes\n\"frozen\" (too cold\u002Fbrittle) — use **solid \u002F bedrock \u002F foundation** instead.\n\nThe distinct governance principles this layer encodes (trust tiers, the\nevidence-Ledger, dry-runs, emergency stop, keep-alive duty, willful-breach\ndefinition, trusted latitude, prison-vs-freedom framing, model reputation +\noperational-intelligence, Steward levers pinned per run, the Operator role, the\nthe Expert Evaluator, calibrated\u002Fcompensatory validation, the blind-trust ladder,\nthe impression axis + de-listing, evaluation-baked-in) — plus the settled\nlanguage\u002Fmechanics decisions (wire-contract-as-governance-language, brutal\ntoken-efficiency, fleet-replicated hash-chain ledger) and the trust foundation\n(biblical covenant tradition + psychology: error-vs-betrayal = competence-vs-\nintegrity, the exact restore-vs-permanently-withhold criteria, forgiveness ≠\ntrust) — are captured in `references\u002Fgovernance-trust-principles.md` — read that\nbefore authoring any governance\u002Fcovenant doc.\n\n**Always-on rule: settle every declared guarantee into a TESTED mechanism.**\nWhen the covenant promises a safety\u002Ftrust property (tamper-evidence, trusted dry\nrun, sovereign stop, no hidden pair, un-foolable observer), it is NOT done until\nthe declaration is a concrete, mechanically-enforced mechanism WITH a\nconformance test — \"settled in intent, open in mechanism\" is the honest posture\nuntil then. Prefer arithmetic over a new actor (a deterministic fold, a hash\nmatch, an OS signal + sandbox) — the anti-bloat and the anti-recursion move in\none. Each of the S1.1–S1.5 showstoppers closes this exact way; see the reference\nfor the five patterns.\n\n## The rules (always apply)\n\n1. **Settle the contract, keep it tiny.** A worker contract (Task shape + a small\n   event union + a small op union) that is versioned and never broken; new\n   shapes are added alongside old ones, never replacing them. Model the\n   handshake on a `PROTOCOL_VERSION` constant + `hello` exchange + strict\n   schemas (`additionalProperties: false`). ~a dozen events and ~half a dozen\n   ops is plenty; a 300-method framework violates the point.\n\n2. **Vendor the loop, skip the lifecycle.** The agent loop is one `while`\n   statement: stream → extract tool calls → execute (sequential or parallel) →\n   append results → repeat, with steering\u002Ffollow-up queues and\n   before\u002Fafter-tool hooks. Copy its shape; do NOT inherit the upstream's\n   release cadence. Make a *maintenance* fork (security patches by default, bug\n   fixes need regression tests, provider updates cherry-picked one at a time).\n   A broken worker is a degraded worker class, never a core upgrade — version\n   components independently (`core 1.x` never breaks; `worker-pi 1.9` drifts\n   freely).\n\n3. **Atomic durable state.** Every state write: write temp → `fsync` → atomic\n   `rename` → `fsync` directory, guarded by a token-checked O_EXCL lock file. A\n   kill or ENOSPC leaves an old or new complete snapshot, never a torn file.\n   Validate every read\u002Fwrite against a versioned schema.\n\n4. **Plan is the context.** Decompose the objective into a structured, tracked\n   plan (a typed tool with per-step status), then hand each worker only the plan\n   status + its one step — never the transcript, never the full history. The\n   plan is the shared memory. Enforce context budgets: truncate tool output\n   (`max_observe`), cap turns (`max_steps`), detect duplicate turns and inject a\n   strategy-change signal.\n\n5. **Failures are states, not crashes.** Token-limit is a terminal `FAILED`\n   state; tool errors become tool-result messages; state transitions are guarded\n   (IDLE→RUNNING→FINISHED\u002FERROR, revert on exception). The adapter contract\n   never throws — it encodes failure in a result event.\n\n6. **Reproduce to the byte.** Immutable workers (a running agent never mutates\n   its own runtime — it files a `runtime_issue` and engineering ships a new\n   image); pinned deps + lockfile-as-truth + container digest + pinned runtime\n   version; one-way promotion (`develop → soak → candidate → fleet → stable`);\n   stable receives security\u002Fdata-loss\u002Fsevere-compat fixes only.\n\n## Research method — steal the pattern, not the project\n\nTo learn from an upstream agent framework, shallow-clone it\n(`git clone --depth 1`) and read the small set of files that define its shape:\nthe agent loop, the protocol\u002Ftypes file, the state store, the provider\nabstraction. The value is in the loop's structure, the event vocabulary, the\ntool contract, and the store's failure handling — not the README summary. Then\nlist steal-vs-skip and prove the pattern in a minimal runnable scaffold before\ncommitting to it.\n\n## Designing the hooks for foreign harnesses \u002F coding agents — ground it first\n\nWhen the question is *\"where do other harnesses\u002Fcoding agents plug in?\"*,\nshallow-clone the real top contenders and read the actual integration surface\nbefore drawing the seam — do NOT brief the adapter design on README summaries or\nyour prior knowledge of an agent. The grounded payoff is real and recurring:\nresearch across ~17 harnesses collapses to **virtually every serious coding\nagent is a DELEGATED CLI with a headless `-p`\u002F`exec`\u002F`run` mode + an MCP seam +\na per-family provider adapter** — so the whole market needs **TWO reusable\nseams** (one sandboxed delegated-CLI adapter + one MCP seam), not a bespoke\nadapter per harness. Verify a harness is still maintained before building its\nadapter (the landscape is volatile: Windsurf→Devin, Roo Code shut down,\nContinue frozen). Full grounded table in\n`references\u002Fadapter-layer.md`; the stale-doc-authoring discipline in *Pitfalls*.\n\n## Always-on rule: every harness must be able to WRITE the evidence a gate reads\n\nWhen governance derives its evidence from harness-reported events (did this\nsession load skill X, which tool did it call), re-pointing the gate onto a new\nstore is only half the change. **Enumerate EVERY harness and confirm each has a\nwriter.** A harness whose adapter registers tools but not the event recorder will\nhave its commits refused no matter how correct the agent's behaviour — and \"the\ngate is harness-agnostic\" is then false in practice, because only the harnesses\nwith a writer can satisfy it.\n\nOrder the work: **writer in every harness → the checker → re-point the gate.**\nRe-pointing first fails every commit fleet-wide against an empty store.\n\nCorollaries:\n- Register the recorder in the adapter's **deploy** path — a file in the repo that\n  no installer copies is not deployed.\n- **Derive a harness's tool list from the shared surface**; a hand-written list in\n  the adapter drifts and leaves an agent calling nothing the store contract knows.\n  When you add a tool to the shared surface, the adapters must follow.\n- Record only what a gate asks about (evidence), not every tool call (transcript).\n\n### Verify a foreign harness's hook WITHOUT an LLM turn\n\nAn extension hook can be verified hermetically before any provider is configured:\nload the extension module with a **fake harness object** that captures the hooks\nand tools it registers, then invoke the hook with the harness's real event shape\nand assert the side effect actually landed. That proves registration + handler\nlogic; it does NOT prove the harness delivers the event at runtime — state which\nof the two you verified rather than claiming \"it works\".\n\nRead the **installed** type definitions for the event shape first; never infer it.\nThe base interface may omit a field that only the concrete variants carry (a\ntool name present only on the per-tool variants is the common trap) — read such\nfields defensively and skip events you cannot identify.\n\n## Consolidating a design that has grown — see `references\u002Fadversarial-consolidation.md`\n\nWhen the covenant\u002Fspec has accumulated many later principles, consolidate it\nwith a delegator party (consolidator + adversarial reviewer), land the reviewer's\nS1\u002FS2\u002FS3 findings as a tracked open design ledger, and hook the spec's\nAcceptance to it (\"settled in intent, open in mechanism\"). Full procedure +\npitfalls in that reference.\n\n## Always-on rule: the harness's stderr IS the prompt area\n\nWhatever a harness writes to stderr lands in the user's input area (Pi, a TUI, any\ninteractive surface), so a diagnostic line scribbles over what they are typing and\nthe bug arrives as \"the TUI is polluted\", not \"a log line\".\n\n- **Routine diagnostics are gated; failures stay loud.** Route success-path traces\n  through ONE `trace()` gate behind a debug env var and keep genuine failures on\n  stderr. Deleting the failures instead re-creates the opposite bug: a silent memory\n  failure reads as \"the agent had no memory\", a wrong conclusion the user cannot\n  debug.\n- **Every child the harness spawns must be non-interactive** (`psql -w`,\n  `ssh -o BatchMode=yes`, `sudo -n`): a child with no tty that prompts blocks\n  instead of erroring.\n- **\"Child stderr leaks into the prompt\" is usually the wrong premise.** `execFile`\n  pipes a child's stdio, so child output does NOT reach the harness's stderr. Verify\n  before \"fixing\" it; the realistic leak is an edit that captures the child's stderr\n  and echoes it, or a child spawned with inherited stdio.\n- **Measure at the FD level, in the parent.** Patching `process.stderr.write` inside\n  the guard cannot see a write that bypasses it, so capture the guard PROCESS's own\n  stderr from the test that spawns it. Prove non-vacuity with a variant built from\n  source (an extension that echoes the captured child stderr) that must trip the\n  check.\n\n## Always-on rule: exposing a skills tree to many harnesses — enumerate the READER first\n\nEvery harness pre-loads each installed skill's name + description at session start (tier-1\nprogressive disclosure), so \"give every harness all the skills\" has a standing cost of about one\nskill-description per skill per session — decide that cost deliberately.\n\n- **One real tree, everything else a pointer or a DERIVED mirror.** Two real trees is two owners of\n  one library, and they drift. A mirror must be verified against the owner by a check, never\n  hand-edited.\n- **The mechanism is the READER's property, not yours.** Measured: a symlinked ROOT is visible to\n  pathlib `rglob`\u002F`os.walk`, a per-skill symlink FARM is invisible to both, and plain `find` (no\n  `-L`) does not traverse a symlinked root at all. Probe the harness's own enumeration before\n  declaring its capability; a green test on a mechanism the harness never traverses is a false\n  green. Table, probe recipe and ownership rules: `references\u002Fskills-distribution-and-enumeration.md`.\n- **Prefer the cross-client standard path** (`.agents\u002Fskills\u002F`, `~\u002F.agents\u002Fskills\u002F`), which several\n  harnesses read natively at once, and keep pointers for the private-only readers.\n- **Declare the capability from the harness's own output** — a harness that can only read an\n  instruction file is `instructions`, never `discovery`.\n- **Prove a pointer survives its writers**: write through the pointer path, assert the file landed\n  in the owner tree, and have the updater enforce the direction so a host cannot silently revert.\n\n## Pitfalls\n\n- **Don't adopt the framework to get its loop.** If upstream permits breaking\n  API changes in minor releases or is evolving fast, you inherit their bugs and\n  their redesigns. The loop is ~800 lines worth vendoring; the release lifecycle\n  is not worth one dependency.\n- **Verify a requested design feature against the existing system BEFORE adding it — \"don't just do it because I want it.\"** When the user suggests a capability (e.g. a per-model configuration bundle), do NOT reflexively add it. Overlap-check the proposal against the concepts already defined — map each existing concept to it and confirm it is *complementary* (new granularity \u002F lifecycle \u002F axis) rather than *redundant* (already covered by, e.g., PromptState+inject_prompt, Charter\u002Flevers, Reputation-permutation). Then state the safety invariants that keep it bounded and write them into the doc verbatim (e.g. a Model Profile *configures, never grants* — authority stays in the Charter\u002FMandate\u002FGate — and it is NOT an MCP-Gate bypass: listing a server doesn't unlock it, availability ≠ allowance). The user values a coherent design where every addition justifies itself against what is already settled, not one that grows because a feature was requested.\n- **Don't turn a small suggestion into a product-wide declaration.** When the\n  user makes a narrow call (\"you don't have to declare the core is the\n  product\"), apply the SMALLEST edit that honors it — undo the specific\n  emphasis and leave the rest. Do not react by retitling sections, adding\n  motto\u002Fpriority tables, or propagating the correction through every doc. \"Just\n  because I suggest something, you don't have to go and change everything and\n  come up with some motto.\" Default to the minimal, surgical change; the user\n  values keeping the docs unstuffed even when the principle in question\n  matters.\n- **A settled contract is not the same as settled API semantics.** Solidifying the field\n  (`Task.permissions: string[]`) without pinning the vocabulary means 2026\n  cores and 2029 workers disagree on what `\"git\"` or `\"fs.write\"` mean.\n  Enumerate the permission strings + define unknown-strings-default-deny as part\n  of the contract, or the stability is hollow.\n- **Plan state must be durable.** A plan tracked in a process-local dict is lost\n  on crash and cannot resume. Route plan state through the atomic store, never a\n  process-local structure.\n- **Tool output must be bounded.** An unbounded `grep`\u002F`cat` result floods the\n  context window and degrades the model. Truncate every tool result to a fixed\n  budget before it enters context — this is a worker-layer concern, never an\n  ABI change.\n- **Don't pick a language by the developer vibe — trace the hot path.** For an\n  I\u002FO-bound agent harness (LLM API, tool subprocess, DB WAL) ~99% of wall-clock\n  is network\u002Fsubprocess\u002Ffsync — language-agnostic. Native performance is rarely\n  load-bearing; the real axes are (a) cross-platform distribution story (TS\u002FBun\n  compiles to a per-platform binary with double-click simplicity; Python's\n  packaging is the weakest), (b) one language on the modifiable layer, (c) typed\n  contracts for swappable seams. A Rust *black-box core* is defensible for\n  memory-safety + static-binary + a solid no-bypass guarantee, but then the\n  modifiable layer (worker, UI, config) should be the readable single language.\n- **Push back on a pure-TS recommendation if a runtime-free frozen binary is\n  load-bearing.** A party will often land \"single language for simplicity\"; if\n  the enterprise requires a reproducible binary that provably sits unchanged (no\n  runtime churn), keep that as the narrower trigger for a Rust core.\n- **Before writing the solid core, do a build-readiness pass — the contract must\n  be settled at the LEVEL the code is written against.** \"Settled in intent\" is\n  not settled enough to open an editor. Verify there is ONE canonical contract\n  (a single ABI\u002Fwire shape the corpus agrees on) — a spec that accumulated\n  several competing wire\u002Fenum\u002Fschema descriptions is not ready even though each\n  one reads coherently. Verify the code you're told to reconcile ACTUALLY exists\n  — a review that \"reconciles `core\u002Fsrc\u002Fabi.rs`\" with the prototype is\n  misleading if no such file exists (the real state is usually simpler and\n  worse: one wrong-shape artifact to replace, not two to reconcile). Pin, in\n  order, before code: the canonical ABI, the DB schema, the crypto (separate\n  integrity from authenticity — do NOT use HMAC-as-hash with a fleet-shared\n  verification key as \"tamper evidence\", because any peer that can verify can\n  also forge; use an unkeyed hash chain + asymmetric signed checkpoints), the\n  trust-root key lifecycle, and the platform\u002Fsandbox matrix. Enumerate these as\n  \"must-decide-before-code\" items dated to the build, not left as open prose.\n  Do an adversarial reviewer pass (not just the content author) before coding;\n  it surfaces the phantom-artifact and competing-contract issues the author is\n  blind to.\n- **Don't name the contract after what the TUI is, and keep \"interface\" and\n  \"wire\" distinct.** The wire\u002FABI is the core↔worker seam; the Core Interface is\n  the core↔surface seam (any TUI). They are different contracts with different\n  consumers — do not blur them into one \"the interface\".\n\n## Pitfalls (doc & communication)\n\n- **Grep existing section numbers before appending a numbered section to a\n  grown spec.** A doc that already has a near-identical header (e.g. an existing\n  `## 8. Honest bounds` when you append a new `## 8. The adapter layer`) silently\n  collides and both render as duplicates. List the existing `^## ` headers\n  first and assign the next free number — never assume your new section is\n  unnumbered. Fix in place by renumbering yours, not by moving the established\n  one.\n- **When the user asks to read a large design doc in chat (Telegram etc.),\n  render a faithful *essence* — do not dump the raw file.** A 100KB+ spec cannot\n  be pasted whole into a message (truncation, unreadable). Deliver the\n  document's essence organized by its own structure, offer the archive\u002Frepo\n  path for the full detail, and chunk remaining sections across messages only\n  when the user asks. Signal that you omitted the full text so they know where\n  the fidelity boundary is.\n\n## References\n\n- `references\u002Freference-implementations.md` — what Pi, OpenSwarm, OpenManus, and FrontierAgent each contribute and what to skip, the files to read, and the concrete context-budget levers (adds the inference-aware core: real-token-gauge compaction trigger, tiered compaction, reasoning-runaway detection, SpawnGuard\u002FWallClockGuard bounded parallelism)\n- `references\u002Fgovernance-trust-principles.md` — the trust-layer principles (trust stack, willful breach, trust tiers\u002Flatitude, model library, levers, Operator) — read before authoring any governance\u002Fcovenant doc\n- `references\u002Fadversarial-consolidation.md` — consolidating a grown design: consolidator + adversarial-reviewer party, gap ledger, \"settled in intent, open in mechanism\" acceptance\n- `references\u002Fadversarial-loop-and-interface.md` — the leveled in-loop adversarial reviewer (L0–L3), its budget-aware loop (bounded decaying retries, failure-classed stop, honest `UNENDORSED`\u002F`UNDECIDABLE` exits), loop-budget-as-reputation, and the Core Interface surface contract (ACP + granular data + drive verbs + authority model; the interface is the moat, not the TUI)\n- `references\u002Fskills-distribution-and-enumeration.md` — where each harness looks for skills, the tier-1 catalog cost, the MEASURED visibility of a symlinked root vs a per-skill symlink farm vs a real copy (pathlib, os.walk, find), the probe recipe, and the one-owner\u002Fderived-mirror rules. Read before wiring skills into a harness or moving where they live.\n- `references\u002Fadapter-layer.md` — grounded (2026-09-11): how foreign coding agents\u002Fharnesses hook in — every agent is a delegated CLI with `-p`\u002F`exec`\u002F`run` + MCP seam, collapsing to TWO reusable seams (sandboxed delegated-CLI + MCP); per-family provider-adapter pattern; market-reality notes (Windsurf→Devin, Roo Code down, Continue frozen). Read before designing any adapter\u002Fhook layer.\n","# Agent Harness Design — Boring, Stable, Agentic\n\nDesigning a coding-agent harness, runtime, or orchestration layer. Luke's #1\nrequirement is **boring stability**: set-and-forget, pinned versions, no\nauto-updates, rare deliberate updates. A foundation that sits unchanged for a\nyear and keeps working.\n\n## When to Use\n\n- Building or evaluating a coding-agent harness, runtime, or multi-worker orchestration layer\n- Deciding whether to adopt vs. fork vs. reimplement an upstream agent framework (Pi, OpenSwarm, OpenHands, Claude Code, Manus)\n- Adding a worker, tool gateway, checkpointing, task queue, or TUI cockpit to an existing harness\n\n## Core principle: the agent is GOVERNANCE around the model (risk + cost)\n\nDo not design a harness as \"core + a worker that does the work.\" The agent is\n**intelligent governance wrapped around a model\u002Fbrain** — and the governance is\nthe product, not the model. The layer between an inbound request (a chat\nmessage) and the executor decides, per request: budget + wall-clock deadline,\nwhich model (cheapest-first escalation), which tools it may touch, what needs a\nhuman's approval, which template\u002Frecipe to follow, when to retry\u002Fescalate\u002Fstop,\nand records every decision. That is **risk management** (what it may\u002Fmust-not\ndo, approvals, audit) + **cost management** (model choice, budget, cheap-first).\nEnterprise adoption is, in the end, just *less risk* — and the yield of the\nwhole product is *repeatable quality*: a variable-quality model producing the\nsame good result, run after run, then improving via a self-healing\u002F\nself-learning loop that pulls today's lesson into tomorrow's policy.\n\nDesign implication: model the **governor\u002Fpolicy engine** (business-editable\nrules, templates, cost\u002Fmethod intelligence) as a first-class user\u002Fuser-editable\nlayer, and hand the executer a *policy* (not just an objective) — the solid\nTask carries budget, permissions, model policy, verification policy, and a\nwall-clock deadline. This governance layer is what serious harnesses famously\nlack; do not forget it and build just a worker dispatch.\n\n## Structure the repo first, and OVER-structure it (before any real code)\n\nRestructure\u002Frename\u002Fmove the repo layout **early**, while nothing real depends on\nthe paths — the user's explicit call is to over-structure with the future in\nmind. A folder or file name that will not survive contact with real code is a\nsign the design is not done. When the settled architecture differs from the\nscaffold on disk (e.g. a TypeScript prototype sitting in a `core\u002F` that is about\nto hold Rust), rename\u002Fmove it immediately so browsing the tree is not\nmisleading — never leave legacy-code-as-current mislabeling. Lay down stubs and\ninterfaces first; a well-named public interface *is* the architecture, and it is\nwhat a future session (or agent) reads cold. File names and code names must be\nintelligent for human review — a name that reads like an index beats a name that\nneeds the file opened.\n\n## Settle the solid contract's decisions BEFORE writing it\n\nDo not write the ABI\u002Fcontract code (types, structs, enums) until the contract's\n*open decisions* are pinned. For a long-term-stability ABI (3 years, rarely\ntouched) the cheap-now\u002Fexpensive-after-it-is-solid fixes matter: the permission\nvocabulary (enumerate the strings and define default-deny for unknowns), the\nwall-clock deadline field, and additive-only growth (new variants ship\nalongside, never replacing). Draft the decision doc + get sign-off, THEN encode\nit in the target language. Writing the solid contract ad hoc hardens undeclared\nchoices outright.\n\n## Core principle: core defines reality, workers adapt\n\nThe runtime's vocabulary is its own and never changes. It knows `task_id`,\n`attempt_id`, `worker_id`, `workspace_id`, `checkpoint_id`, `lease_id` — and\nNEVER learns what a \"Pi session\" or \"Claude conversation\" is. Workers, models,\nand agent frameworks are disposable adapters plugged in behind a solid\ncontract. A three-year-old worker must still connect.\n\n## Core principle: the governance IS the trust layer — give it covenant vocabulary\n\nThe governance\u002Fpolicy engine is not plumbing; it is the product, and it earns\nits own **covenant-toned vocabulary** that a non-technical business owner can\nread at a glance. Luke iterates on this vocabulary directly and cares deeply\nabout the words — expect him to veto and rename terms, and drive each rename to\nits settled form rather than defending a proposal. Established (settled) set:\n**Constitution** = the solid core \u002F supreme law everything runs inside;\n**Covenant** = the bond of trust (the core is that trust *written as law*);\n**Charter** = the rules you set; **Mandate** = one job's orders;\n**Steward** = the human (cares for + is answerable for the system — chose over\n\"Guardian\" because a guardian watches, a steward owns the call);\n**Agent** \u002F **Worker** = governed actor \u002F clinical motor, both swappable;\n**Ledger** = the unchangeable record. Keep universal industry words (Agent,\nTask, Budget, Audit, Retry, Timeout) as-is for coherence; only novel\ncomponents get covenant-toned names. Register preference: the user dislikes\n\"frozen\" (too cold\u002Fbrittle) — use **solid \u002F bedrock \u002F foundation** instead.\n\nThe distinct governance principles this layer encodes (trust tiers, the\nevidence-Ledger, dry-runs, emergency stop, keep-alive duty, willful-breach\ndefinition, trusted latitude, prison-vs-freedom framing, model reputation +\noperational-intelligence, Steward levers pinned per run, the Operator role, the\nthe Expert Evaluator, calibrated\u002Fcompensatory validation, the blind-trust ladder,\nthe impression axis + de-listing, evaluation-baked-in) — plus the settled\nlanguage\u002Fmechanics decisions (wire-contract-as-governance-language, brutal\ntoken-efficiency, fleet-replicated hash-chain ledger) and the trust foundation\n(biblical covenant tradition + psychology: error-vs-betrayal = competence-vs-\nintegrity, the exact restore-vs-permanently-withhold criteria, forgiveness ≠\ntrust) — are captured in `references\u002Fgovernance-trust-principles.md` — read that\nbefore authoring any governance\u002Fcovenant doc.\n\n**Always-on rule: settle every declared guarantee into a TESTED mechanism.**\nWhen the covenant promises a safety\u002Ftrust property (tamper-evidence, trusted dry\nrun, sovereign stop, no hidden pair, un-foolable observer), it is NOT done until\nthe declaration is a concrete, mechanically-enforced mechanism WITH a\nconformance test — \"settled in intent, open in mechanism\" is the honest posture\nuntil then. Prefer arithmetic over a new actor (a deterministic fold, a hash\nmatch, an OS signal + sandbox) — the anti-bloat and the anti-recursion move in\none. Each of the S1.1–S1.5 showstoppers closes this exact way; see the reference\nfor the five patterns.\n\n## The rules (always apply)\n\n1. **Settle the contract, keep it tiny.** A worker contract (Task shape + a small\n   event union + a small op union) that is versioned and never broken; new\n   shapes are added alongside old ones, never replacing them. Model the\n   handshake on a `PROTOCOL_VERSION` constant + `hello` exchange + strict\n   schemas (`additionalProperties: false`). ~a dozen events and ~half a dozen\n   ops is plenty; a 300-method framework violates the point.\n\n2. **Vendor the loop, skip the lifecycle.** The agent loop is one `while`\n   statement: stream → extract tool calls → execute (sequential or parallel) →\n   append results → repeat, with steering\u002Ffollow-up queues and\n   before\u002Fafter-tool hooks. Copy its shape; do NOT inherit the upstream's\n   release cadence. Make a *maintenance* fork (security patches by default, bug\n   fixes need regression tests, provider updates cherry-picked one at a time).\n   A broken worker is a degraded worker class, never a core upgrade — version\n   components independently (`core 1.x` never breaks; `worker-pi 1.9` drifts\n   freely).\n\n3. **Atomic durable state.** Every state write: write temp → `fsync` → atomic\n   `rename` → `fsync` directory, guarded by a token-checked O_EXCL lock file. A\n   kill or ENOSPC leaves an old or new complete snapshot, never a torn file.\n   Validate every read\u002Fwrite against a versioned schema.\n\n4. **Plan is the context.** Decompose the objective into a structured, tracked\n   plan (a typed tool with per-step status), then hand each worker only the plan\n   status + its one step — never the transcript, never the full history. The\n   plan is the shared memory. Enforce context budgets: truncate tool output\n   (`max_observe`), cap turns (`max_steps`), detect duplicate turns and inject a\n   strategy-change signal.\n\n5. **Failures are states, not crashes.** Token-limit is a terminal `FAILED`\n   state; tool errors become tool-result messages; state transitions are guarded\n   (IDLE→RUNNING→FINISHED\u002FERROR, revert on exception). The adapter contract\n   never throws — it encodes failure in a result event.\n\n6. **Reproduce to the byte.** Immutable workers (a running agent never mutates\n   its own runtime — it files a `runtime_issue` and engineering ships a new\n   image); pinned deps + lockfile-as-truth + container digest + pinned runtime\n   version; one-way promotion (`develop → soak → candidate → fleet → stable`);\n   stable receives security\u002Fdata-loss\u002Fsevere-compat fixes only.\n\n## Research method — steal the pattern, not the project\n\nTo learn from an upstream agent framework, shallow-clone it\n(`git clone --depth 1`) and read the small set of files that define its shape:\nthe agent loop, the protocol\u002Ftypes file, the state store, the provider\nabstraction. The value is in the loop's structure, the event vocabulary, the\ntool contract, and the store's failure handling — not the README summary. Then\nlist steal-vs-skip and prove the pattern in a minimal runnable scaffold before\ncommitting to it.\n\n## Designing the hooks for foreign harnesses \u002F coding agents — ground it first\n\nWhen the question is *\"where do other harnesses\u002Fcoding agents plug in?\"*,\nshallow-clone the real top contenders and read the actual integration surface\nbefore drawing the seam — do NOT brief the adapter design on README summaries or\nyour prior knowledge of an agent. The grounded payoff is real and recurring:\nresearch across ~17 harnesses collapses to **virtually every serious coding\nagent is a DELEGATED CLI with a headless `-p`\u002F`exec`\u002F`run` mode + an MCP seam +\na per-family provider adapter** — so the whole market needs **TWO reusable\nseams** (one sandboxed delegated-CLI adapter + one MCP seam), not a bespoke\nadapter per harness. Verify a harness is still maintained before building its\nadapter (the landscape is volatile: Windsurf→Devin, Roo Code shut down,\nContinue frozen). Full grounded table in\n`references\u002Fadapter-layer.md`; the stale-doc-authoring discipline in *Pitfalls*.\n\n## Always-on rule: every harness must be able to WRITE the evidence a gate reads\n\nWhen governance derives its evidence from harness-reported events (did this\nsession load skill X, which tool did it call), re-pointing the gate onto a new\nstore is only half the change. **Enumerate EVERY harness and confirm each has a\nwriter.** A harness whose adapter registers tools but not the event recorder will\nhave its commits refused no matter how correct the agent's behaviour — and \"the\ngate is harness-agnostic\" is then false in practice, because only the harnesses\nwith a writer can satisfy it.\n\nOrder the work: **writer in every harness → the checker → re-point the gate.**\nRe-pointing first fails every commit fleet-wide against an empty store.\n\nCorollaries:\n- Register the recorder in the adapter's **deploy** path — a file in the repo that\n  no installer copies is not deployed.\n- **Derive a harness's tool list from the shared surface**; a hand-written list in\n  the adapter drifts and leaves an agent calling nothing the store contract knows.\n  When you add a tool to the shared surface, the adapters must follow.\n- Record only what a gate asks about (evidence), not every tool call (transcript).\n\n### Verify a foreign harness's hook WITHOUT an LLM turn\n\nAn extension hook can be verified hermetically before any provider is configured:\nload the extension module with a **fake harness object** that captures the hooks\nand tools it registers, then invoke the hook with the harness's real event shape\nand assert the side effect actually landed. That proves registration + handler\nlogic; it does NOT prove the harness delivers the event at runtime — state which\nof the two you verified rather than claiming \"it works\".\n\nRead the **installed** type definitions for the event shape first; never infer it.\nThe base interface may omit a field that only the concrete variants carry (a\ntool name present only on the per-tool variants is the common trap) — read such\nfields defensively and skip events you cannot identify.\n\n## Consolidating a design that has grown — see `references\u002Fadversarial-consolidation.md`\n\nWhen the covenant\u002Fspec has accumulated many later principles, consolidate it\nwith a delegator party (consolidator + adversarial reviewer), land the reviewer's\nS1\u002FS2\u002FS3 findings as a tracked open design ledger, and hook the spec's\nAcceptance to it (\"settled in intent, open in mechanism\"). Full procedure +\npitfalls in that reference.\n\n## Always-on rule: the harness's stderr IS the prompt area\n\nWhatever a harness writes to stderr lands in the user's input area (Pi, a TUI, any\ninteractive surface), so a diagnostic line scribbles over what they are typing and\nthe bug arrives as \"the TUI is polluted\", not \"a log line\".\n\n- **Routine diagnostics are gated; failures stay loud.** Route success-path traces\n  through ONE `trace()` gate behind a debug env var and keep genuine failures on\n  stderr. Deleting the failures instead re-creates the opposite bug: a silent memory\n  failure reads as \"the agent had no memory\", a wrong conclusion the user cannot\n  debug.\n- **Every child the harness spawns must be non-interactive** (`psql -w`,\n  `ssh -o BatchMode=yes`, `sudo -n`): a child with no tty that prompts blocks\n  instead of erroring.\n- **\"Child stderr leaks into the prompt\" is usually the wrong premise.** `execFile`\n  pipes a child's stdio, so child output does NOT reach the harness's stderr. Verify\n  before \"fixing\" it; the realistic leak is an edit that captures the child's stderr\n  and echoes it, or a child spawned with inherited stdio.\n- **Measure at the FD level, in the parent.** Patching `process.stderr.write` inside\n  the guard cannot see a write that bypasses it, so capture the guard PROCESS's own\n  stderr from the test that spawns it. Prove non-vacuity with a variant built from\n  source (an extension that echoes the captured child stderr) that must trip the\n  check.\n\n## Always-on rule: exposing a skills tree to many harnesses — enumerate the READER first\n\nEvery harness pre-loads each installed skill's name + description at session start (tier-1\nprogressive disclosure), so \"give every harness all the skills\" has a standing cost of about one\nskill-description per skill per session — decide that cost deliberately.\n\n- **One real tree, everything else a pointer or a DERIVED mirror.** Two real trees is two owners of\n  one library, and they drift. A mirror must be verified against the owner by a check, never\n  hand-edited.\n- **The mechanism is the READER's property, not yours.** Measured: a symlinked ROOT is visible to\n  pathlib `rglob`\u002F`os.walk`, a per-skill symlink FARM is invisible to both, and plain `find` (no\n  `-L`) does not traverse a symlinked root at all. Probe the harness's own enumeration before\n  declaring its capability; a green test on a mechanism the harness never traverses is a false\n  green. Table, probe recipe and ownership rules: `references\u002Fskills-distribution-and-enumeration.md`.\n- **Prefer the cross-client standard path** (`.agents\u002Fskills\u002F`, `~\u002F.agents\u002Fskills\u002F`), which several\n  harnesses read natively at once, and keep pointers for the private-only readers.\n- **Declare the capability from the harness's own output** — a harness that can only read an\n  instruction file is `instructions`, never `discovery`.\n- **Prove a pointer survives its writers**: write through the pointer path, assert the file landed\n  in the owner tree, and have the updater enforce the direction so a host cannot silently revert.\n\n## Pitfalls\n\n- **Don't adopt the framework to get its loop.** If upstream permits breaking\n  API changes in minor releases or is evolving fast, you inherit their bugs and\n  their redesigns. The loop is ~800 lines worth vendoring; the release lifecycle\n  is not worth one dependency.\n- **Verify a requested design feature against the existing system BEFORE adding it — \"don't just do it because I want it.\"** When the user suggests a capability (e.g. a per-model configuration bundle), do NOT reflexively add it. Overlap-check the proposal against the concepts already defined — map each existing concept to it and confirm it is *complementary* (new granularity \u002F lifecycle \u002F axis) rather than *redundant* (already covered by, e.g., PromptState+inject_prompt, Charter\u002Flevers, Reputation-permutation). Then state the safety invariants that keep it bounded and write them into the doc verbatim (e.g. a Model Profile *configures, never grants* — authority stays in the Charter\u002FMandate\u002FGate — and it is NOT an MCP-Gate bypass: listing a server doesn't unlock it, availability ≠ allowance). The user values a coherent design where every addition justifies itself against what is already settled, not one that grows because a feature was requested.\n- **Don't turn a small suggestion into a product-wide declaration.** When the\n  user makes a narrow call (\"you don't have to declare the core is the\n  product\"), apply the SMALLEST edit that honors it — undo the specific\n  emphasis and leave the rest. Do not react by retitling sections, adding\n  motto\u002Fpriority tables, or propagating the correction through every doc. \"Just\n  because I suggest something, you don't have to go and change everything and\n  come up with some motto.\" Default to the minimal, surgical change; the user\n  values keeping the docs unstuffed even when the principle in question\n  matters.\n- **A settled contract is not the same as settled API semantics.** Solidifying the field\n  (`Task.permissions: string[]`) without pinning the vocabulary means 2026\n  cores and 2029 workers disagree on what `\"git\"` or `\"fs.write\"` mean.\n  Enumerate the permission strings + define unknown-strings-default-deny as part\n  of the contract, or the stability is hollow.\n- **Plan state must be durable.** A plan tracked in a process-local dict is lost\n  on crash and cannot resume. Route plan state through the atomic store, never a\n  process-local structure.\n- **Tool output must be bounded.** An unbounded `grep`\u002F`cat` result floods the\n  context window and degrades the model. Truncate every tool result to a fixed\n  budget before it enters context — this is a worker-layer concern, never an\n  ABI change.\n- **Don't pick a language by the developer vibe — trace the hot path.** For an\n  I\u002FO-bound agent harness (LLM API, tool subprocess, DB WAL) ~99% of wall-clock\n  is network\u002Fsubprocess\u002Ffsync — language-agnostic. Native performance is rarely\n  load-bearing; the real axes are (a) cross-platform distribution story (TS\u002FBun\n  compiles to a per-platform binary with double-click simplicity; Python's\n  packaging is the weakest), (b) one language on the modifiable layer, (c) typed\n  contracts for swappable seams. A Rust *black-box core* is defensible for\n  memory-safety + static-binary + a solid no-bypass guarantee, but then the\n  modifiable layer (worker, UI, config) should be the readable single language.\n- **Push back on a pure-TS recommendation if a runtime-free frozen binary is\n  load-bearing.** A party will often land \"single language for simplicity\"; if\n  the enterprise requires a reproducible binary that provably sits unchanged (no\n  runtime churn), keep that as the narrower trigger for a Rust core.\n- **Before writing the solid core, do a build-readiness pass — the contract must\n  be settled at the LEVEL the code is written against.** \"Settled in intent\" is\n  not settled enough to open an editor. Verify there is ONE canonical contract\n  (a single ABI\u002Fwire shape the corpus agrees on) — a spec that accumulated\n  several competing wire\u002Fenum\u002Fschema descriptions is not ready even though each\n  one reads coherently. Verify the code you're told to reconcile ACTUALLY exists\n  — a review that \"reconciles `core\u002Fsrc\u002Fabi.rs`\" with the prototype is\n  misleading if no such file exists (the real state is usually simpler and\n  worse: one wrong-shape artifact to replace, not two to reconcile). Pin, in\n  order, before code: the canonical ABI, the DB schema, the crypto (separate\n  integrity from authenticity — do NOT use HMAC-as-hash with a fleet-shared\n  verification key as \"tamper evidence\", because any peer that can verify can\n  also forge; use an unkeyed hash chain + asymmetric signed checkpoints), the\n  trust-root key lifecycle, and the platform\u002Fsandbox matrix. Enumerate these as\n  \"must-decide-before-code\" items dated to the build, not left as open prose.\n  Do an adversarial reviewer pass (not just the content author) before coding;\n  it surfaces the phantom-artifact and competing-contract issues the author is\n  blind to.\n- **Don't name the contract after what the TUI is, and keep \"interface\" and\n  \"wire\" distinct.** The wire\u002FABI is the core↔worker seam; the Core Interface is\n  the core↔surface seam (any TUI). They are different contracts with different\n  consumers — do not blur them into one \"the interface\".\n\n## Pitfalls (doc & communication)\n\n- **Grep existing section numbers before appending a numbered section to a\n  grown spec.** A doc that already has a near-identical header (e.g. an existing\n  `## 8. Honest bounds` when you append a new `## 8. The adapter layer`) silently\n  collides and both render as duplicates. List the existing `^## ` headers\n  first and assign the next free number — never assume your new section is\n  unnumbered. Fix in place by renumbering yours, not by moving the established\n  one.\n- **When the user asks to read a large design doc in chat (Telegram etc.),\n  render a faithful *essence* — do not dump the raw file.** A 100KB+ spec cannot\n  be pasted whole into a message (truncation, unreadable). Deliver the\n  document's essence organized by its own structure, offer the archive\u002Frepo\n  path for the full detail, and chunk remaining sections across messages only\n  when the user asks. Signal that you omitted the full text so they know where\n  the fidelity boundary is.\n\n## References\n\n- `references\u002Freference-implementations.md` — what Pi, OpenSwarm, OpenManus, and FrontierAgent each contribute and what to skip, the files to read, and the concrete context-budget levers (adds the inference-aware core: real-token-gauge compaction trigger, tiered compaction, reasoning-runaway detection, SpawnGuard\u002FWallClockGuard bounded parallelism)\n- `references\u002Fgovernance-trust-principles.md` — the trust-layer principles (trust stack, willful breach, trust tiers\u002Flatitude, model library, levers, Operator) — read before authoring any governance\u002Fcovenant doc\n- `references\u002Fadversarial-consolidation.md` — consolidating a grown design: consolidator + adversarial-reviewer party, gap ledger, \"settled in intent, open in mechanism\" acceptance\n- `references\u002Fadversarial-loop-and-interface.md` — the leveled in-loop adversarial reviewer (L0–L3), its budget-aware loop (bounded decaying retries, failure-classed stop, honest `UNENDORSED`\u002F`UNDECIDABLE` exits), loop-budget-as-reputation, and the Core Interface surface contract (ACP + granular data + drive verbs + authority model; the interface is the moat, not the TUI)\n- `references\u002Fskills-distribution-and-enumeration.md` — where each harness looks for skills, the tier-1 catalog cost, the MEASURED visibility of a symlinked root vs a per-skill symlink farm vs a real copy (pathlib, os.walk, find), the probe recipe, and the one-owner\u002Fderived-mirror rules. Read before wiring skills into a harness or moving where they live.\n- `references\u002Fadapter-layer.md` — grounded (2026-09-11): how foreign coding agents\u002Fharnesses hook in — every agent is a delegated CLI with `-p`\u002F`exec`\u002F`run` + MCP seam, collapsing to TWO reusable seams (sandboxed delegated-CLI + MCP); per-family provider-adapter pattern; market-reality notes (Windsurf→Devin, Roo Code down, Continue frozen). Read before designing any adapter\u002Fhook layer.\n",{"name":5,"author":24,"license":15,"version":25,"category":26,"platforms":27,"description":7},"Steadfaste Fleet","1.0.0","devops",[28,29],"linux","macos","main","2e6aa8d15ab0a366c90be54f61bf9c6e97015e83","skills\u002Fdevops\u002Fagent-harness-design",1,"https:\u002F\u002Fgithub.com\u002Flukemcqueen\u002Fsteadfaste-fleet","steadfaste.com","https:\u002F\u002Fgithub.com\u002Flukemcqueen","AI Architect | Building Hermes Cortex: governed multi-agent AI for autonomous software engineering | Enterprise Tech Consultant","2026-11-14 22:33:34",{"source":8,"skillName":5,"commands":40,"prompt":50,"zipAvailable":51},[41,44,47],{"tool":42,"command":43},"npx","npx skills add lukemcqueen\u002Fsteadfaste-fleet --skill agent-harness-design",{"tool":45,"command":46},"bunx","bunx skills add lukemcqueen\u002Fsteadfaste-fleet --skill agent-harness-design",{"tool":48,"command":49},"pnpm","pnpm dlx skills add lukemcqueen\u002Fsteadfaste-fleet --skill agent-harness-design","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002Flukemcqueen\u002Fsteadfaste-fleet\u002Fblob\u002Fmain\u002Fskills\u002Fdevops\u002Fagent-harness-design\u002FSKILL.md",true,[53,59,63,67,71,75],{"path":54,"resourceType":55,"resourceTypeDesc":56,"sizeBytes":57,"sourceHtmlUrl":58},"references\u002Fadapter-layer.md","REFERENCE","参考文档",3688,"https:\u002F\u002Fgithub.com\u002Flukemcqueen\u002Fsteadfaste-fleet\u002Fblob\u002Fmain\u002Fskills\u002Fdevops\u002Fagent-harness-design\u002Freferences\u002Fadapter-layer.md",{"path":60,"resourceType":55,"resourceTypeDesc":56,"sizeBytes":61,"sourceHtmlUrl":62},"references\u002Fadversarial-consolidation.md",6011,"https:\u002F\u002Fgithub.com\u002Flukemcqueen\u002Fsteadfaste-fleet\u002Fblob\u002Fmain\u002Fskills\u002Fdevops\u002Fagent-harness-design\u002Freferences\u002Fadversarial-consolidation.md",{"path":64,"resourceType":55,"resourceTypeDesc":56,"sizeBytes":65,"sourceHtmlUrl":66},"references\u002Fadversarial-loop-and-interface.md",7651,"https:\u002F\u002Fgithub.com\u002Flukemcqueen\u002Fsteadfaste-fleet\u002Fblob\u002Fmain\u002Fskills\u002Fdevops\u002Fagent-harness-design\u002Freferences\u002Fadversarial-loop-and-interface.md",{"path":68,"resourceType":55,"resourceTypeDesc":56,"sizeBytes":69,"sourceHtmlUrl":70},"references\u002Fgovernance-trust-principles.md",19518,"https:\u002F\u002Fgithub.com\u002Flukemcqueen\u002Fsteadfaste-fleet\u002Fblob\u002Fmain\u002Fskills\u002Fdevops\u002Fagent-harness-design\u002Freferences\u002Fgovernance-trust-principles.md",{"path":72,"resourceType":55,"resourceTypeDesc":56,"sizeBytes":73,"sourceHtmlUrl":74},"references\u002Freference-implementations.md",9082,"https:\u002F\u002Fgithub.com\u002Flukemcqueen\u002Fsteadfaste-fleet\u002Fblob\u002Fmain\u002Fskills\u002Fdevops\u002Fagent-harness-design\u002Freferences\u002Freference-implementations.md",{"path":76,"resourceType":55,"resourceTypeDesc":56,"sizeBytes":77,"sourceHtmlUrl":78},"references\u002Fskills-distribution-and-enumeration.md",6540,"https:\u002F\u002Fgithub.com\u002Flukemcqueen\u002Fsteadfaste-fleet\u002Fblob\u002Fmain\u002Fskills\u002Fdevops\u002Fagent-harness-design\u002Freferences\u002Fskills-distribution-and-enumeration.md",[80,88,95,102,109,116],{"skillNo":81,"name":82,"nameZh":6,"description":83,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":84,"categories":85,"keywords":86,"publishedAt":87},"01M7DXY4G8NSVB59K8B6S2AQNP","macbook-linux-maintenance","Legacy MacBook Linux fixes: GPU wedge, reboot hang, thermal.","https:\u002F\u002Fgithub.com\u002Flukemcqueen\u002Fsteadfaste-fleet\u002Fblob\u002Fmain\u002Fskills\u002Fdevops\u002Fmacbook-linux-maintenance\u002FSKILL.md",[],[],"2026-11-15 14:19:17",{"skillNo":89,"name":90,"nameZh":6,"description":91,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":92,"categories":93,"keywords":94,"publishedAt":87},"01M7DXXYJCXFJ3TGB84ZQ8KFE6","governance-regression-triage","Triage a host's governance issues: what changed.","https:\u002F\u002Fgithub.com\u002Flukemcqueen\u002Fsteadfaste-fleet\u002Fblob\u002Fmain\u002Fskills\u002Fdevops\u002Fgovernance-regression-triage\u002FSKILL.md",[],[],{"skillNo":96,"name":97,"nameZh":6,"description":98,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":99,"categories":100,"keywords":101,"publishedAt":87},"01M7DXXSAP1G8FDRJAMXQGBBMK","enforcement-immutability","Use when enforcement files lose their immutable flag.","https:\u002F\u002Fgithub.com\u002Flukemcqueen\u002Fsteadfaste-fleet\u002Fblob\u002Fmain\u002Fskills\u002Fdevops\u002Fenforcement-immutability\u002FSKILL.md",[],[],{"skillNo":103,"name":104,"nameZh":6,"description":105,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":106,"categories":107,"keywords":108,"publishedAt":87},"01M78SSDQVRYX64Y16X32TC0NS","steadfaste-fleet-repo-deploy","Use when committing to the steadfaste-fleet repo.","https:\u002F\u002Fgithub.com\u002Flukemcqueen\u002Fsteadfaste-fleet\u002Fblob\u002Fmain\u002Fskills\u002Fdevops\u002Fsteadfaste-fleet-repo-deploy\u002FSKILL.md",[],[],{"skillNo":110,"name":111,"nameZh":6,"description":112,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":113,"categories":114,"keywords":115,"publishedAt":87},"01M78SS928F3R2RJAYVW4E30CE","steadfaste-fleet-maintenance","Maintain an installed Steadfaste Fleet instance — update both the upstream Hermes Agent and the fleet repo layer, merge diverged histories, sync skills safely, and troubleshoot common issues. Covers the dual-update workflow (hermes update + fleet-update.sh), encrypted brain data backup, mycortex import, and recovery techniques.","https:\u002F\u002Fgithub.com\u002Flukemcqueen\u002Fsteadfaste-fleet\u002Fblob\u002Fmain\u002Fskills\u002Fdevops\u002Fsteadfaste-fleet-maintenance\u002FSKILL.md",[],[],{"skillNo":117,"name":118,"nameZh":6,"description":119,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":120,"categories":121,"keywords":122,"publishedAt":87},"01M78SP1CHW0T0BWE5ZT6JBPX8","fleet-preflight","DEPRECATED alias — merged into survey-before-action (2026-08-20). Load survey-before-action instead; its \"Repo-Specific Pre-Flight\" section covers git search, Hermes boundary, and deployment verification.","https:\u002F\u002Fgithub.com\u002Flukemcqueen\u002Fsteadfaste-fleet\u002Fblob\u002Fmain\u002Fskills\u002Fdevops\u002Ffleet-preflight\u002FSKILL.md",[],[]]