Skip to content

About

Agent configuration for agentic-job runs: per-agent settings and AGENTS.md

Resources

Stars

0 stars

Watchers

0 watching

Forks

 
 

Repository files navigation

agenthome

This text is LLM generated.

Agent configuration for agentic-job runs: per-agent settings and AGENTS.md

Tooling

Configuration is edited directly in the checked-in files; no renderer or build step is needed. Run just check-config (or python3 tests/check_config.py) with Python 3.9 or later to lint the configuration and run its positive and negative fixtures. This uses only the standard library, without Cargo, downloads, package installation or network schema fetching.

This repository has no compiled application or Rust toolchain requirement. Offline eval summarization is provided by agent-board eval summarize MANIFEST RUNS; see agent-board's eval documentation. No JavaScript is left, and just check fails if a .js, .mjs or .cjs file comes back.

Navigation tools

Callers provision tools through their own setup steps or images and pass an explicit PATH before launching agents. Nothing in this repository installs tools on an agent's behalf. When ast-outline and ast-grep are absent, use rg and scoped reads; do not retry or install missing tools.

The Check workflow demonstrates ordinary Linux x86_64 caller setup in a job-local prefix, without privileges or credentials. It retains reviewed release SHA-256 pins for ast-grep 0.50.0, uv 0.13.1 and cgwalters-forge-misctool 0.1.0, rejects checksum mismatches before extraction or execution, and bounds downloads and provisioning. ast-outline 1.9.0 uses uv tool install --no-build, the full constraints/ast-outline.txt and --exclude-newer 2026-10-21T00:00:00Z. No sdist build runs; uv ignores hashes in constraints, so wheels remain trusted as PyPI served them before that cutoff. This is not a package sandbox.

just check runs just check-config and offline model discovery tests. CI separately provisions tools and checks their exact versions and structural navigation. No generic installer or manifest interface remains. Do not restore agentic-job's rejected [setup.tools] mechanism. External caller package settings still need their agenthome-only Rust requirements removed; this repository no longer builds or tests a Rust crate.

Installer caller audit (2026-11-14)

Refreshed the earlier inventory against agenthome 4e68032821d38fa0f49601fd4882c68f4afd3bbe and current default branches. The only active check-install caller was this repository's Check workflow; it now owns setup directly. The Justfile wrappers, CLI variants, installer-only tests and installation instructions were the other local install-tools references; they are removed in this slice.

Tracker at 595cc08ed43c4c5245059ffe11265172b9a0d7f7 has only commented setup.tools examples in .github/agentic-job/hosted.toml:70–73 and acp-experiment.toml:71–74, not active calls. Their agenthome revision 9b43eb81e1266ec762a6c49a4e31fb6b4ff3e0e0 is a historical pin, not evidence of a current deployment. Agentic-job at df6134e1009b8774c264de440028a4e324e43711 refuses setup.tools (in its configuration parser); its docs/dispatch.md:376–393 discusses the former interface, not an active call. Agent-board at 157fff2c117c6ea1ea59af2b8517963baa806d43 has no active agenthome installer reference. Their workflows' Cargo invocations build their own programs, not agenthome. Configuration/profile consumers do not require the removed CLI. No active external dependency was found in those default branches; other branches and deployed revisions remain unverified.

PR #87's draft and human-approval hold are unchanged. This slice neither runs paid evals/hosted role probes nor changes canonical roles or model routing.

The root AGENTS.md guides work on this repository only. Its Code navigation paragraph is copied into opencode/AGENTS.md (the file deployed to opencode's config directory, and as opencode/CLAUDE.md to Claude's), and just check keeps the two copies the same. agentic-job runs that turn on agent-skills with this repository as their configuration repository receive skills/, skills/ast-explore included.

The ast-explore skill is copied from cgwalters-bot/homegit, dotfiles/.agents/skills/ast-explore/SKILL.md, at source commit c027d9f4c28c8e21df37f82e978b3fc55fe73cab.

opencode configuration

Model discovery

just discover-models previews newer stable models in the families already declared by the praxis prompt tiers; just discover-models --write writes the bump, including role profile references and regression-test IDs. Preview is the default so inspecting releases cannot accidentally change routing. Set BROKER_BASE_URL (including /v1) and BROKER_API_KEY in the caller environment. Nothing schedules this by default. Discovery checks models.dev (opencode's catalog), OpenAI's Markdown model overview and per-model pages, and Anthropic's model overview. It never adds a tier, changes a fallback, or enables Terra.

Public releases are not evidence of broker availability: only IDs from the broker's /models, or completed Responses probes with matching model IDs when that route returns 404/405, are eligible. There are at most eight probes per invocation, each requesting 16 output tokens at low effort; a broker may ignore that token cap. Requests have a 60-second socket timeout and a 16 MiB response bound. The Linux CLI also has a ten-minute elapsed-time deadline so a slow stream cannot keep discovery running indefinitely. Authenticated redirects are refused so a broker login redirect cannot leak the run token. Discovery is intentionally separate from the network-free just check.

tools/model-sources.json records provider prices and limits for each bump, the URLs corroborating them, and configured caps. Existing local context/input and output caps are retained, not automatically expanded. Cache creation keeps the input rate because the broker meters it as plain input; published provider cache-write rates are recorded separately. OpenAI numeric facts must agree with models.dev or discovery fails without writing. Claude families are supported only if already declared; their numeric facts come from models.dev, with model IDs checked against Anthropic's overview. No Claude pool is declared today. Unparseable sources, missing credentials or ambiguous metadata fail the job; they are not treated as a successful no-change check.

just discover-models --fixtures tests/fixtures/model-discovery.json --write replays the first bump offline. The fixture contains selected verbatim source excerpts and catalog fields recorded on 2026-11-17, plus sanitized successful broker probe outcomes; it contains no endpoint, token or response IDs. Replaying against the bumped tree is a no-op. The weekly workflow must be installed by the controller from the accompanying PR description; it is not enabled here.

opencode/ holds opencode.json (the one provider and its models) and opencode-runner.json (the unattended-run profile merged over it) for agentic-job runs: set the run's agent-config-path to opencode, and its configuration repository and ref through the caller's own variables. The provider address and key are placeholders a run replaces; just check-config enforces the placeholders, since this repository is public.

opencode-local.json is an opt-in profile for interactive sessions on a self-hosted OpenAI-compatible endpoint (llama.cpp router models), with a trimmed local agent for small GPUs. Merge it over the deployed configuration with OPENCODE_CONFIG=~/.config/opencode/opencode-local.json opencode (tested with opencode 1.18.35, using its provider/npm/options format; native V2 providers entries are ignored by that launcher). It enables only its own provider and routes every agent, title and compaction included, to it; just check-config refuses a profile that leaves any agent on praxis. Its endpoint is a placeholder too, replaced in the operator's private derived configuration at provider.local.options.baseURL. Before starting a session, preview model resolution without inference or applying caller changes: OPENCODE_CONFIG=~/.config/opencode/opencode-local.json opencode --pure models local. The output must list the three local/ model IDs, not just exit successfully: 1.18.35 also exits zero for Provider not found. This is a configuration preview, not proof that the endpoint serves those models or supports tool calls. Project configuration loads after it, so a repository's own opencode.json can still re-enable praxis.

The runner follows the review procedure in the skill named by a task, such as agentic-job's implement skill and its diff-reviewer subagent. The build prompt in opencode-runner.json defers to that procedure and keeps its one read-only general reviewer for tasks whose named skills don't define a review. That prompt differs from homegit's copy; keep the change on a re-sync.

opencode.json and opencode-runner.json are copied from homegit dotfiles/.config/opencode/ at source commit c027d9f4c28c8e21df37f82e978b3fc55fe73cab, with the provider address replaced by a placeholder.

opencode/AGENTS.md holds the standing rules; the procedures are skills in skills/, loaded when a task needs them. Both are cut down from agenthome#6 (opencode/AGENTS.md, which became the rules and the commit-messages, upstream-text and subagent-review skills) and from homegit's dotfiles/.agents/skills/commit-review/SKILL.md at the commit above. They don't name an operator or bot login or a homegit tool. just check enforces this. agenthome#8 is closed without deciding how a run learns the operator and bot logins, so the skills that need the operator (commit-messages, upstream-text, commit-review) cannot know it yet. There is no issue for that follow-up yet.

The rules are written for opencode runs and Claude runs. opencode runs load AGENTS.md; the agentic-job adapter for Claude loads only CLAUDE.md. opencode/CLAUDE.md is a copy that just check keeps byte-identical, so edit opencode/AGENTS.md and copy it over with cp opencode/AGENTS.md opencode/CLAUDE.md. Both need the run's agent-config-path set to opencode. A Claude run gets the rules only once agentic-job dispatch forwards the configuration repository for Claude, as it does for opencode and Vibe. Vibe runs do not get the rules: they take only models from config.toml.

Proposed prompt-tier source of truth

prompt-tiers.json is the proposed agenthome-owned source for per-pool prompt tiers and role assignments. These are prompt capability/cost labels, not tracker's pool failover tiers: resolution never switches pools. The proposed names and mapping are not an operator-approved model table. Only praxis is represented, using models already configured in opencode/opencode.json; no other pool's support has been verified. terra is explicitly unavailable (model: null) and falls back to sol. Each entry declares its fallback, with null marking the end of the path. An unavailable entry follows that edge; a declared model must exist in the tool catalogue. Role assignments preserve the currently working model choices, including architect using sol, rather than implicitly promoting it to astra. Model identifiers belong in this data or tool configuration, not role prompts.

just check-config (tests/check_config.py) prevents a misspelled selection from choosing an unconfigured model and a fallback cycle from hanging resolution. It checks every edge, including those behind available models, and walks each path at most once per tier. Unknown pools, tiers, undeclared models, cycles and paths without terminal models fail. The role assignments must resolve to the models opencode/opencode.json and opencode/opencode-runner.json already configure, so the manifest describes the current configuration rather than changing it. This is offline configuration validation, not broker enforcement or a model-availability probe. Nothing reads the manifest at run time; tiered hosted runs and an agentic-job#1367 consumer are not wired up.

Shared role agents

The tiered chain has five roles: explore, plan, execute-lead, worker and review, named role-explore and so on. Each is written twice, and the two are edited together: as a Claude Code Markdown agent in roles/claude/agents/, and as an opencode agent entry in roles/opencode/eval-tiered.json, an opt-in profile for the tiered arm of the evaluation below that makes role-execute-lead the default agent. Edit these files directly. just check-config requires the same description and prompt in both, each role's access level and prompt tier as tests/check_config.py pins them, the opencode model that tier resolves to in prompt-tiers.json's praxis pool, no model or tier named in a prompt, and the formatting the files have now (the profile as JSON.stringify(value, null, 2)).

The standalone tests retain the invariant test names from the former Rust configuration suites. Their mutation groups cover all 32 role/tier fixtures, 9 provider and 3 runner fixtures, and 25 standing-rule fixtures, in their original order. Eval serialization tests now belong to agent-board; the eval digest format is unchanged.

The prefix keeps the roles from replacing opencode's built-in plan and explore, Claude Code's own subagents, or agentic-job's embedded diff-reviewer; the lead and the reviewer follow any review procedure defined by a skill named in the task. Explore, plan and review cannot edit; only the lead can delegate, so a worker cannot start a chain of its own. The read-only roles keep the shell for the navigation tools, as the runner's explore does, so their prompts rather than a tool rule keep them read-only. Claude Code agents take model: inherit: the manifest has no verified Claude pool, and Claude Code subagents cannot start subagents, so the lead works there only as the main agent. opencode/opencode-runner.json keeps its direct-worker policy.

The supported opencode source-subagent delivery is an inline opencode-runner.json profile read from the configuration repository's selected path. It does not require a root agents/ installer. The pinned source audit establishes that contract, not actual role/model execution in a deployed runner. The closed installer issue is not installation proof. Opt-in complete bundles for all three arms can be generated through the existing configuration resolver; see the protocol. No ordinary caller or runner setting changes, and measurement is still blocked on capacity, enforcement and execution evidence.

Offline role evaluation

eval/README.md defines the opt-in three-arm historical evaluation protocol and the input format of agent-board eval summarize MANIFEST RUNS, documented in agent-board. The sample still needs authorized sizing evidence and a runner that installs the role agents. No research runs or workflow changes are enabled. eval/summary.json remains the recorded pending output; synthetic eval tests run in agent-board, not just check here.

Tracker supervisor profile

skills/supervisor holds the hosted supervisor's judgment brief. .github/agentic-job/profiles/supervisor.md is the tracker profile template; copy it to that same path in tracker, enable agent-skills with agenthome as the configuration repository, and select it from the trusted supervisor caller. It doesn't enable a workflow here. The skill's request examples come from agent-board's docs/supervisor-requests.md at commit 9d36c74f5b8a71f94d9a95583f473591f37ec001; tests/fixtures/agent-board-supervisor-actions.jsonl vendors that example block, and just check fails when the skill differs, so a sync updates the commit, the fixture and the skill together. Before enabling apply, prove a preview run lists the requests and doesn't apply any, including the final state-body replacement; that proof belongs to tracker#687.

About

Agent configuration for agentic-job runs: per-agent settings and AGENTS.md

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages