| name | tangent |
| description | ML experiment toolkit and autonomous agent. Say "tangent" for help, "tangent <subagent>" for a specific tool, or "tangent auto" for the full autonomous experiment loop. Requires the published Tangle CLI (`uv tool install tangle-cli` or `uvx --from tangle-cli tangle ...`) and a reachable Tangle backend. |
| allowed-tools | ["Bash","Read","Write","Glob","Grep","Agent","dispatch"] |
Tangent
Thin index. Every file linked here is the source of truth for its topic — read it
when you need it, don't paraphrase from this file.
For the CLI surface, invocation rule, auth, artifacts, learnings corpus, logs, and
annotations, the single source of truth is
OSS-CONVENTIONS.md. When any skill file (this one included)
needs a command, a flag, or an env var, it cites that document rather than
re-deriving it.
First, every session
Skills live in-repo (checked into the tangle-cli repo) — there is no fetch or
bundle-refresh step. Relative cross-references in this skill (agents/*.md,
references/*.md) resolve directly on disk. Read
references/setup.md to set up the CLI and auth.
Normal usage runs the published CLI: install persistently with uv tool install tangle-cli and use tangle …, or run one-off commands with uvx --from tangle-cli tangle …. If intentionally validating a local tangle-cli checkout, prefix examples with uv run. See OSS-CONVENTIONS.md §1.
Commands
tangent — print the help block below.tangent <subagent> — read agents/<subagent>.md and follow it.tangent auto — run the autonomous loop. See Auto Mode.
Subagents: Agent file:
tangent debugger agents/debugger.md
tangent researcher agents/researcher.md
tangent reporter agents/reporter.md
tangent reviewer agents/reviewer.md
tangent builder agents/builder.md
tangent auth agents/auth-wizard.md
tangent new-scenario agents/scenario-builder.md
Automation:
tangent auto — Run full autonomous 8-step experiment loop
References (read on demand)
Tools
Always use the tangle CLI via Bash. Do not use any tangle MCP tools.
Run tangle quickstart to discover commands. See
references/tangle-tools.md.
Cancel a run: tangle sdk pipeline-runs cancel RUN_ID
Background execution: dispatch. Subagents: agents/*.md.
Scenarios
A scenario is any repo with a Tangle pipeline. Contents: scenario.yaml, MEMORY.md,
skills/, sessions/, logs/, pipeline.yaml. No scenario? tangent new-scenario.
For inspiration: references/example-scenarios/INDEX.md.
Auto Mode
Autonomous MLE agent that iterates on an ML pipeline. Tunes parameters, selects
features, changes data, analyzes results, modifies pipelines.
Memory & run_id
- Long-term:
MEMORY.md — best config, lessons, session index. < 3000 tokens.
- Short-term:
sessions/YYYY-MM-DD.md — daily log, append-only.
- Active runs: written to
MEMORY.md on Step 3 submission, cleared on Step 5 completion.
run_id is a first-class session concept. Every Tangle pipeline-run has one;
the session log tracks them in order, and learnings recorded to the corpus are
keyed by <scenario>/<run_id>. Always record the run_id returned by
tangle sdk pipeline-runs submit. See the learnings corpus in
references/knowledge-corpus.md and
OSS-CONVENTIONS.md §6.
Procedure
Read each step file only when you reach that step — do not pre-read all step files.
Execute the step, verify the gate checklist, do not proceed until every gate passes.
Each step gate ends with a "reload skills + context" checkbox — actually re-read
the step file and active agent files at that point, don't rely on stale memory.
Loop: Step 7 → Step 1 if not converged. Read references/tangle-tools.md and
references/event-log.md at Step 0 only.
Extracting metrics
value = metrics
for key in path.split("."):
value = value[int(key) if key.isdigit() else key]
