thrashr888/shift/shift-dev

Working on Shift

Layout

Build before you test

make build

bin/shift-agent and the tests load compiled modules from build/ and fall back to source when a `.go` is stale. The fallback is silent, so an edit you forgot to compile does not fail — it just tests the old module in some paths and the new one in others. Run make build after editing any .scm under src/ or extensions/. make -n build lists what is stale without compiling, and the DEV pane shows the same thing.

Tests

make test          # every Guile suite plus every Python test; ~3 minutes
make check         # the same target

make test runs longer than a default command timeout, so give it room or run it in the background rather than assuming it hung.

One suite at a time:

guile -L src -L extensions -C build test/run.scm test/receipt.scm

This asks for approval every time and is meant to. test/run.scm loads whatever path follows it, so allowlisting any prefix that reaches it would let a file be written and then executed without approval — the shell boundary this project keeps closed. make build, make test, make check and make -n are allowlisted instead.

Registering a new module or suite

• A new src/live-agent/NAME.scm is compiled automatically: the Makefile globs

CORE_SOURCES.

• A new test/NAME.scm is not run automatically. Add NAME to the suite

list in the Makefile's test: target or the suite never runs, and a green make test will mean nothing about it.

• A new test/NAME.py likewise needs its own python3 test/NAME.py line.

Two traps worth knowing

SRFI-64 tests after `test-end` are silently skipped. Appending a case below (test-end "name") leaves it unexecuted while the suite still reports success, with the pass count unchanged. If a case you just added does not appear in the output, check that it sits above test-end.

Tool schemas are charged on every request. Several integration tests pin a tight context-limit, so lengthening a tool description in src/live-agent/tools.scm can push an unrelated turn over its budget and fail it with "Context still exceeds the budget". Put instructions the model needs only after calling a tool in that tool's output rather than its description.

Debugging a session

traces.jsonl in the session folder holds every span. In a session, the traces tool searches it and fetches a span by span_id; outside one, python3 scripts/evals.py session NAME gives a per-turn review with rounds, failures, waits and judge disagreements.