0.1.7

Published to PyPI: https://pypi.org/project/grapharc/0.1.7/

pip install --upgrade grapharc

Two defects closed, both found by re-verifying a stale bug backlog against main rather than from a report.

admission accepted a sentinel pointing the wrong way (#108)

START is the graph's entry and END its exit, but the endpoint check tested only whether an endpoint was a sentinel, never which side of the edge it sat on. END -> x and x -> START were admitted and then died in Materializer.

The run doesn't proceed either way — what was wrong is where the failure landed. GovernedLoop charges a MaterializationError to max_consecutive_execution_failures (2) rather than max_consecutive_rejections (3), so a planner got fewer retries for a mistake admission is supposed to catch than for one it does catch, and two in a row ended the run as EXECUTION_FAILED — a stop reason claiming the graph ran when nothing had. It also reached the planner as prose scraped off an exception rather than a code and a remedy, with nothing on the admission event's failed-check list, because admission hadn't failed.

a dead sandbox child escaped as a bare EOFError (#111)

poll() is true when the pipe is readable, and a closed pipe is readable — so a child killed rather than raising (os._exit, a segfault, an OOM-kill) fell through the timeout guard into recv(). AgentNode catches it under its blanket except Exception and renders f"TOOL_ERROR: {exc}", and str(EOFError('')) is '' — so the model was handed TOOL_ERROR: with nothing after the colon, and the trace recorded a failure it couldn't explain either.

Now a RuntimeError naming the tool and the exit code, with a negative one rendered as the signal that killed it. Deliberately still not a SandboxViolation: a child dying is not evidence it tried to escape confinement.

Full detail in CHANGELOG.md.

0.1.6

Published to PyPI: https://pypi.org/project/grapharc/0.1.6/

pip install --upgrade grapharc

What ships

• grapharc[mcp] was broken on arrival (#101). mcp>=1.2 resolved to 2.0, where mcp.server.fastmcp no longer exists, so a fresh install could not import the server. uv.lock pinned 1.28 and hid it from every local test; only a resolver starting from pyproject.toml — a real user — ever saw it. Now mcp>=1.2,<2, with grapharc/mcp/__init__.py resolving the SDK-backed names lazily so the subpackage imports without the extra.
• grapharc go <run-dir> silently re-executed an already-executed plan (#100). Bare go always skipped executed plans; the named-directory form did not, so re-issuing it ran the whole graph again at exit 0 and overwrote the stamp. Two fixes: the record accumulates executed_run_ids instead of clobbering a scalar, and — the one that matters — a plan carrying an executed_run_id is now refused with exit 2 before anything runs, unless --again says so. An approval binds to a fingerprint that does not change between runs, so this was one human yes authorising N executions of a mutating plan.
• The deep dive's Verified this pass figures rotted, twice (#42). They are now what one command re-derives, and tests/test_deep_dive.py fails the line rather than letting it drift — including a guard that an unowned figure appearing on the line is itself a failure.
• Five max_seconds timing tests failed only on a loaded machine (#40). The deadline machinery was never what failed; the assertions were. Every wall-clock bound now sits between the deadline and the uninterrupted body, verified 20/20 under three parallel suites at load average 33-35.
• Novita joins the model gateway (#102) — thanks @jax-novita. A fifth OpenAI-wire backend behind a novita extra, probed by grapharc models --check like the others.

Behaviour change

A second grapharc go <run-dir> on an executed plan now exits 2 instead of 0. That is the fix rather than a side effect; --again restores the old behaviour explicitly.

Full detail for every entry is in CHANGELOG.md.