Own the harness's behaviour here, in harness/ - #8
Open
TomProkop wants to merge 2 commits into
Open
Conversation
Adds `agent/`: how an agent harness should behave in a TALXIS workspace — session-wide instructions and the session-start briefing — declared in `agent/instructions.json`, which names its payload files so this repo owns its own layout. This is the process/know-how half of the setup. The environment half — which tools a box has, and installing or updating them — stays in TALXIS/tools-agentbox, whose provisioner fetches these files over HTTPS and applies each to whichever harness the box runs: the instructions become managed memory or user-level custom instructions, the briefing becomes a session-start hook. Deliberately not a plugin component. `/plugin install` doesn't deliver it, nothing under plugins/ is touched, and there's no version bump — a directory under plugins/ would be classified as a plugin by validate.mjs and forced to carry a marketplace entry and a semver version. So it sits at the root, next to scripts/. scripts/validate.mjs grows an agent/ block, since nothing outside plugins/ was checked before and a typo here breaks provisioning for every new box: the manifest parses, the files it declares exist and are non-empty, stray undeclared .md files are rejected, and the payloads carry the same host-agnostic and zero-toolchain-prose rules the skills do — installs and version checks are txc doctor's job, and what a box has installed isn't something to assert in a prompt. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01KaKGKCwRPFoHcZkRJZxbn2
TomProkop
pushed a commit
to TALXIS/tools-agentbox
that referenced
this pull request
Aug 28, 2026
Splits the config along the line between the box and the agent: this repo owns the environment, TALXIS/skills owns process and know-how. src/agent/agent.json keeps what provisioning a box actually means — which plugin marketplaces and plugins it gets — and gains an "instructions" pointer at the behaviour config, which now lives in TALXIS/skills alongside the Skills it belongs to (TALXIS/skills#8). The system prompt and initial message leave this repo. configure-agent-harness.sh resolves in two stages. download_config already fetched a manifest and resolved payload filenames against the URL curl landed on, which is exactly what the second stage needs, so it splits into download_manifest (the bootstrap) and download_instructions (the behaviour manifest plus its payloads). Both stages fail loudly with the URL named: a box with a plugin list but no house rules is half-configured, so an unreachable instructions manifest is as fatal as an unreachable bootstrap one. AGENTBOX_INSTRUCTIONS_URL / AGENTBOX_INSTRUCTIONS_DIR mirror the existing AGENTBOX_CONFIG_* pair. Tests now drive a fixture in test/agent/fixture/ rather than the real config, so CI covers the mechanism without reaching the network or depending on another repo — which it otherwise would have to, now that half the config lives elsewhere. Adds a case for unreadable instructions alongside the existing one for an unreadable bootstrap manifest. Unchanged: src/container/claude-code/session-start-hook.json and the README step for it. Keeping txc and the Dataverse templates current is tool updating, so it stays an agentbox concern rather than moving into the plugin. Requires the talxis.com/agentbox-instructions short link, and TALXIS/skills#8 merged first. Verified end-to-end against that PR's branch: behaviour fetched over the network, applied, and reported present in a live Claude session's context. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01KaKGKCwRPFoHcZkRJZxbn2
"Agent" is already overloaded in this repo — AGENTS.md, agent skills, coding agents — and `agents/` is a plugin component directory in both harnesses (subagent definitions), so a root `agent/` read as adjacent to something it isn't. This directory configures the harness, which is what the docs already called it. Directory, manifest path, validation messages and both house-rule files renamed together; no behaviour change. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01KaKGKCwRPFoHcZkRJZxbn2
TomProkop
changed the title
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Adds a top-level
harness/directory holding how an agent harness should behave in a TALXIS workspace: the session-wide instructions and the session-start briefing, declared inharness/instructions.json(which names its payload files, so this repo owns its own layout).Companion to TALXIS/tools-agentbox#35, which is updated to fetch these files instead of carrying its own copies.
harness/rather thanagent/: "agent" is already overloaded here —AGENTS.md, agent skills, coding agents — andagents/is a plugin component directory in both harnesses (subagent definitions), so a rootagent/would read as adjacent to something it isn't. This directory configures the harness.The boundary
This repo says how an agent should work. Which tools a box has, and installing or updating them, stays in
tools-agentbox. Its provisioner fetches these files over HTTPS at startup and applies each to whichever harness the box runs — the instructions become managed memory or user-level custom instructions, the briefing becomes a session-start hook.Deliberately not a plugin component
/plugin installdoes not deliver this, nothing underplugins/is touched, and there is no version bump. A directory underplugins/would be classified as a plugin byvalidate.mjsand forced to carry a marketplace entry, the Agent Plugins$schemaand a semver version — so this sits at the root, next toscripts/. Both existing scripts only walkplugins/, soharness/is inert to them, and thevalidateworkflow has no path filter, so a PR touching onlyharness/still produces a green required check.Validation
scripts/validate.mjsgrows aharness/block, because nothing outsideplugins/was checked before and a typo here breaks provisioning for every new box:harness/instructions.jsonparses, and declares at least one payload.mdfilename that exists and is non-empty.mdfile inharness/is an error~/.claude-style paths — this text reaches every harness), and zero toolchain prose (no installs or updates; that istxc doctor's job, T1)Verified by breaking each rule in turn and confirming the specific error, then confirming a clean tree passes both gates:
node scripts/generate-manifests.mjs --check && node scripts/validate.mjs.One thing worth a reviewer's eye
The prompt this replaces (drafted in the agentbox PR) opened with a toolchain inventory — "
txc,pac,dotnetare installed, neverapt-getthem". That is exactly the prose this repo keeps out of skills, and by the environment/know-how split it is agentbox's business, so it is not carried over. What moved is the behaviour: ask the CLI instead of guessing, credentials come from Valet and are never handled directly, source control is the record, production is opt-in, report outcomes honestly. If you want the environment facts asserted to the agent somewhere, agentbox is the place for them.🤖 Generated with Claude Code
https://claude.ai/code/session_01KaKGKCwRPFoHcZkRJZxbn2