hcaiano/skills
Persistent two-agent pairing across claude, codex, cursor, grok, and opencode, in a Herdr tab or headless outside it.
Related neighbors and high-traction skills in the same topics — useful to compare before installing.
Helps users discover and install agent skills when they ask questions like "how do I do X", "fi…
3.3M installsBrowser automation CLI for AI agents. Use when the user needs to interact with websites, includ…
810.4K installsReview UI code for Web Interface Guidelines compliance. Use when asked to "review my UI", "chec…
617.3K installsBuild, deploy, evaluate, optimize, fine-tune, and manage Microsoft Foundry agents, models, and …
576.5K installsPrepare azd-based Azure projects for deployment: generates azure.yaml, infrastructure (Bicep/Te…
568.3K installsOther skills from hcaiano/skills · top by installs.
npx skills add hcaiano/skills
Declared targets from SKILL.md / docs. Unmarked agents are not listed — the skill may still install via the CLI.
main
Parsed from SKILL.md frontmatter.
Files included with this skill beyond the listing page.
SKILL.md
7,190 B
SUMMARY.md
335 B
Pair two agents on one task: you are the lead, and the partner is one of claude, codex, cursor, grok, or opencode — any of them except the CLI you are already running, because two panes of one CLI echo rather than review. Keep the pair and its sid alive across tasks, accepted work cycles, and context compaction, naming its sid in every command — one lead can run several pairs at once. Keep protocol headers and identifiers literal.
The refusal is based only on the CLI kind. Two different harnesses may use the same underlying model or provider and still form a valid pair.
Two backends carry the same protocol. Inside Herdr the partner is a visible pane the user can read and interject in; outside it the partner is a persistent headless session of the partner CLI.
Look for an existing pair before proposing one: a session recorded for this Herdr tab, or a headless one in this repository (the backend reference names the exact command). Before a new work cycle, read the measured pool state as defined in [references/models.md](references/models.md). Resume an existing pair when its pool is available. When its pool is protected, state its use, pace, and reset, then ask whether to spend it or end the session and choose another partner. When its partner, model, or effort differs from what the user just asked for, say so and keep going — respawning discards the pair's whole history, and a model is changed by ending the pair, not by restarting its pane.
With no pair to resume, ask the user in plain chat text — no structured-question tool needed — for three things, and start nothing until they answer. Read pool state first and do not recommend a protected or unavailable pool. The user may explicitly choose a protected pool after you state its use, pace, and reset:
CLI default, or any model name they name. When choosing amodel, effort, or pool, read [references/models.md](references/models.md); it owns the risk, speed, context, catalog, and effort rubric. For a cursor partner run cursor-agent --list-models; for Grok run grok models; for OpenCode run opencode models; for the others take the user's answer as given.
(--effort low|medium|high|xhigh|max), grok (--reasoning-effort), codex (modelreasoningeffort), or cursor (an [effort=…] suffix on the model name, so it needs a model too), or OpenCode (--variant on headless opencode run; its Herdr TUI does not expose this flag).
Then ask for the role, which sets who holds the write leases by default:
peer (default): equals. Split scopes, one lease per scope, review eachother's ready.
executor: the partner holds the write leases and implements; you plan andreview. Any individual task still redistributes leases.
The backend records partner, model, effort, and role in the session, so a resumed pair keeps them without asking again.
Read one backend reference in full and follow it:
HERDR_ENV=1 → [Herdr backend](references/herdr.md). Its preconditionsown the herdr CLI check: a Herdr environment missing the CLI stops there instead of falling through to a hidden headless session.
An inbound [agent ...] or [herdr-pair control ...] line always means the Herdr backend, whatever else the environment looks like.
Messages start with a header:
[agent <from> -> <to> kind=<kind> sid=<sid>]
<body>The Herdr backend adds a [herdr-pair control seq=<n>: ...] line under it.
Use these kinds:
task: propose or update the work split and its write leases. Begin amid-flight stop with STOP — <reason>.
review: request review with file paths and a short change summary.question: ask for clarification before proceeding.ready: report changed files, validation, and residual risk.accepted: accept the partner's ready and advance the work cycle.blocked: name the user decision required to continue.stalemate: report the same disagreement repeated twice without movement.handoff: return control to the user in normal local output.Send every partner message through the backend's helper. Reserve normal assistant output for a header-free user handoff.
Give one agent the write lease for each file scope: owner, target files, forbidden changes, validation, and stop point. The partner stays read-only on that scope until handoff.
The role sets the opening distribution, and each task may redistribute from there. Under executor the partner holds every lease until a task says otherwise, and you review rather than edit.
Under peer the partners are equals: propose a scope split — one write lease per scope, each partner implementing its own scopes and reviewing the other's ready — and include enough context for independent work; the partner accepts or counters before implementing. A task with no independent scopes takes one lease: the holder drives while the partner reviews, and the lease alternates across tasks.
Announce every deliberate broken-checkout window. Before mutation testing, bisect, or a deliberate revert, send a task naming the affected paths and stop condition; send ready after the tree is restored and verified. Do not wait for accepted: the notices are the protection. A partner that has seen an open window treats that checkout's test results as unusable until the close notice arrives, and asks rather than reports. Run an experiment that cannot be announced, or that lasts more than a few minutes, in a separate worktree.
Continue while producing useful artifacts. Five consecutive turns with no new code, test result, decision, or narrowed option require a handoff. Reset the count on real progress. Settle a factual disagreement with one direct proof or focused test before it can become a stalemate. Send stalemate after the same judgment call repeats twice without movement.
Two accepted statuses complete one work cycle. The initiator gives the user a local handoff naming the result, verification evidence, unresolved issues, and every pair pane, worktree, or watcher still active. Both agents may idle; the next task resumes the same pair and sid. The session remains active.
Session renewal is a user decision. If a fresh pair is chosen, carry continuity in a written checkpoint in its first task.
blocked and stalemate also hand off without deleting the session. End the session only when the user explicitly asks to end the pair; the backend reference owns that command.