glm-5.3-flash
fb7cfe3851
Target 4 from the fuzzing plan (the alkcall-bug-class hunt): a #[derive(Arbitrary)] op sequence drives the producer session pump (drive_session_pre_negotiated) over a duplex pair against a fuzz-local mock backend. - fuzz/shared/src/session_opseq.rs: op vocabulary (client writes incl. invalid-type/oversize-length barrier probes, backend production/EOF, exit resolve/fail/drop, control messages, bounded reads, yields), kill-guard exit future (ADR-005 observability), delivery model, and the invariants: kill-on-Drop (kill_fired == !exit_resolved), exit chunk is last with first-issued code -1 on wait-failure (ADR-004), lossless per-stream FIFO + single stdout sentinel, stdin prefix-losslessness (post-shutdown chunks never delivered), control dispatch bounds, take-once allocation, teardown termination. - Harness-model correction caught by a 60s smoke run before the campaign (the alkcall §7.7 pattern, third time): a failed client write can mean the session completed and dropped the server duplex half while the write was in flight - the harness now drains and requires the exit chunk on any write/shutdown error (premature close without it is a finding). No adapter code changed. - 12 committed seeds via the deterministic generator, hand-encoded against the arbitrary 1.4.x derive layout; the encoding is pinned by a seed-decode test. Existing chunk_frame/negotiation_frame/ control_json seed corpora are byte-identical to before. - fuzz/.gitignore: the per-target seed negations never matched (corpus/* excludes the parent dir; git cannot re-include beneath an excluded dir). Fixed with the !corpus/*/ + corpus/*/* + !corpus/*/seed-* recipe - new seed files were silently ignored until now. - Campaign: 10 min detached via run-detached.sh, exited 0, artifact dir empty - 42.3k execs across 33 fork jobs, 0 crash/oom/timeout, cov 3934 -> 4075 edges. Grown corpus excised per the corpus policy. Verification: corpus replay green on stable (9 tests), cargo test 113 + --all-features 156, clippy stable/wasm/fuzz-shared, fmt, doc, publish dry-run, wasm check - all pass.
alktty fuzzing
cargo-fuzz targets for the wire-facing parse surfaces. The design and
operating rules live in docs/plans/fuzzing.md (adopted from alkcall's
docs/research/fuzzing.md); this README is the operational cheat-sheet.
Layout
fuzz_targets/— nightly-onlyfuzz_target!binaries (thin wrappers).shared/— stable-toolchain library holding the invariant logic; the corpus replay tests run here on plaincargo test.corpus/<target>/— committed seeds (regenerate withpython3 fuzz/gen_fuzz_seeds.py).artifacts/— gitignored crash/oom/timeout artifacts + campaign logs.
Targets
| Target | Drives |
|---|---|
chunk_frame |
ChunkReader/ChunkWriter (5-byte chunk codec, ADR-001) |
negotiation_frame |
NegotiationReader/NegotiationWriter + NegotiateRequest + error_response_bytes + the cross-codec peek-disambiguation seam |
control_json |
ControlMessage::from_slice/to_json + signal_from_name |
session_opseq |
the producer session pump (drive_session_pre_negotiated) + a fuzz-local mock backend — the stateful op-sequence target (ADR-004/ADR-005 invariants; the alkcall-bug-class hunt) |
Running a campaign — always detached
Agent sessions must never run fuzzing in the foreground (an OOM in a target can take down the session host; see docs/plans/fuzzing.md §3). Use the detached runner:
fuzz/run-detached.sh chunk_frame
# poll:
tail -n 50 fuzz/artifacts/chunk_frame-*.log
ls fuzz/artifacts/chunk_frame/
pgrep -f "cargo fuzz run chunk_frame"
FUZZ_RUNTIME_SECS=1800 fuzz/run-detached.sh negotiation_frame for a
longer campaign. The runner pins -fork=1 -rss_limit_mb=2048 -malloc_limit_mb=2048 -timeout=25 and detaches via setsid + nohup.
Corpus replay (the standing fuzz gate)
cargo test --manifest-path fuzz/shared/Cargo.toml
Replays every committed seed through the same invariant functions the fuzzer runs, on stable, no nightly needed. Part of the release verification checklist (AGENTS.md).
Toolchain notes
rust-toolchain.toml pins nightly for this subtree only (llvm-tools
required by cargo-fuzz); the crate itself stays on stable at MSRV 1.88.
cargo fuzz build must be run with the fuzz dir as CWD or trust the
toolchain file (rustup resolves per directory).
Run campaigns before releases and after touching src/wire.rs,
src/negotiation.rs, src/control.rs, the adapter, or the session pump.
session_opseq notes
The op-sequence target is inherently slower than the byte-format targets
(each input drives a full tokio session with bounded reads and yields —
tens of exec/s per worker, not millions). That is expected: the value is
state-space exploration, not throughput. It uses tokio::time::timeout
internally, so the nightly target needs the --features time tokio
surface already present via the shared crate's deps — no extra flags.
Its invariants (kill-on-Drop, exit-chunk-is-last, sentinel/FIFO
semantics) are the ones unit tests assert individually; the fuzzer
explores interleavings no example test can enumerate.