Skip to content

AGENT_INSTALL.md

Latest commit

 

History

History
144 lines (106 loc) · 4.89 KB

AGENT_INSTALL.md

File metadata and controls

144 lines (106 loc) · 4.89 KB
Raw

Agent Install

BioProspector is a local agent skill. The install pattern is the same across harnesses: symlink the skills/bioprospector/ directory into the harness's skills folder, then describe the campaign to your agent.

Codex Skill

From the repository root:

mkdir -p "$HOME/.codex/skills"
ln -sfn "$PWD/skills/bioprospector" "$HOME/.codex/skills/bioprospector"

Claude Code Skill

Claude Code reads skills from ~/.claude/skills/. From the repository root:

mkdir -p "$HOME/.claude/skills"
ln -sfn "$PWD/skills/bioprospector" "$HOME/.claude/skills/bioprospector"

Claude Code discovers the skill by its SKILL.md frontmatter. Once linked, ask Claude Code to use the bioprospector skill and it will follow the campaign loop documented there.

Symphony Workers

Symphony workers can use BioProspector by pointing the worker's repo root at this checkout and referencing the skill in the worker's task brief. The agent brief CLI produces a Symphony-compatible kickoff packet:

python3 skills/bioprospector/scripts/bioprospector_agent_brief.py \
  --campaign skills/bioprospector/examples/huperzine-frontier-public-v0/campaign-manifest.json \
  --out .runtime/symphony-brief/huperzine \
  --prefix HUPERZINE \
  --profile public-demo \
  --mode goal \
  --agent codex

The generated agent-goal-prompt.txt, agent-brief.md, commands.sh, and agent-brief.json can be loaded into a Symphony task or piped into a Linear-driven worker queue. The --agent flag accepts codex, claude, or goal for different orchestrator conventions.

Other Agents

Any agent that can read repository files works the same way. Point it at the checkout, ask it to read skills/bioprospector/SKILL.md, README.md, and the docs index, then have it write outputs under .runtime/. The skill's CLIs run with stock Python 3.11+ and no third-party dependencies.

Quick Sanity Check

After install, ask the agent to run the doctor and report what it sees:

Use the bioprospector skill. Run the local doctor and the public audit,
then tell me which optional tools are available and which example campaigns
are ready to inspect.

The agent should call python3 skills/bioprospector/scripts/bioprospector_doctor.py and python3 scripts/public_audit.py . and summarize the result.

Then ask the agent to use the bioprospector skill. The skill will direct the agent to the route-expansion, enzyme-frontier, work-graph, provider-readiness, dossier, example, and release-check scripts in this checkout.

Copy One Prompt

Use the bioprospector skill in this checkout. Turn <target molecule> in <host>
into a route atlas, enzyme frontier, candidate package, Pareto shortlist,
review dossier, and Linear-ready issue graph under .runtime/. Keep everything
local for now and do not write raw sequences, credentials, provider identifiers,
signed URLs, private paths, or large databases into the repo.

Create a BioProspector campaign for <target molecule> in <host>. Start from
templates/target-contract.example.json, generate a scaffold under
.runtime/first-campaign/, run preflight and input audit, then tell me the route
families, likely dark steps, candidate-mining lanes, and smallest set of
operator decisions still needed before expansion.

More task prompts are in AGENT_PLAYBOOK.md.

Repository Guidance

For agents that read repository instructions, include this short boundary in the target repo's AGENTS.md:

Use the BioProspector skill for bioprospecting, pathway planning, enzyme
candidate review, provider readiness, or claim-audited dossier work.

Do not store API keys, private sequences, raw reads, model weights, database
mirrors, provider pod IDs, signed URLs, or unpublished collaborator data in the
repo. Keep large and private artifacts outside git and reference them only by
safe pointers, checksums, or accession IDs.

CLI Setup

The public package is source-checkout oriented so scripts, schemas, examples, and agent skill files stay in one place.

python3 -m pip install -e .
bioprospector --help

If you invoke console scripts from outside the checkout, point them back to this repo:

BIOPROSPECTOR_REPO_ROOT=/path/to/bioprospector \
  bioprospector doctor --json

An installed wheel without BIOPROSPECTOR_REPO_ROOT is intentionally limited: version/help commands work, but script-backed commands fail closed rather than guessing a checkout from the current directory.

What Agents Should Do First

Agents should run local checks before generating provider plans:

python3 skills/bioprospector/scripts/bioprospector_doctor.py
python3 scripts/public_audit.py .

For a campaign, run preflight and input audit before asking operator questions. Ask only the smallest reversible set of questions needed for planning.

For ready-to-use prompts and closeout expectations, continue with AGENT_PLAYBOOK.md.