Fission

fission-automation

Canonical automation runner for Fission quality pipelines.

This crate is used to:

• run repeatable nir-check lanes over sentinel binaries
• run source-oracle semantic quality checks through source-semantic-check
• emit machine-readable artifacts (summary.json, decision_insights.json, candidate lists)
• compare current runs against a baseline
• provide a Go/Stop gate for follow-up recovery work (for example P5H3G readiness)

---

Quick start

From repository root:

cargo run -p fission-automation -- nir-check --lane nir

By default, outputs are written under:

• benchmark/artifacts/automation/<lane>-<profile>-<run_id>/

and latest is mirrored to:

• benchmark/artifacts/automation/latest/<lane>/

Source semantic smoke check:

cargo run -p fission-automation -- source-semantic-check \
  --no-build \
  --fission-bin target/release/fission_cli \
  --function-name fibonacci \
  --jobs 1 \
  --timeout-sec 30

This wraps benchmark/source_semantic_benchmark/run_source_semantic_benchmark.py. The Python runner outputs stay under source_semantic/; automation-level summary.json, summary.md, decision_insights.json, and source_semantic_rows.json are written at the run root.

---

Important options

• --lane <name>
  • canonical lane is nir
  • preview is accepted but deprecated alias
• --run-profile {fast|mid|full}
  • fast: short feedback loop (smaller effective limits/timeouts)
  • mid: default balanced mode
  • full: broader validation window
• --manifest <path>
  • override the default lane manifest at crates/fission-automation/config/sentinel_sets.toml
• --focus-top-mismatch <N>
  • uses baseline candidates to focus run targets to binaries implicated by top conditional_tail_exit_mismatch rows
  • ideal for rapid iteration on conditional-tail recovery patches
• --baseline <path/to/summary.json>
  • enables baseline delta + row-level mismatch diff + go/stop gating
• --functions-limit <N> and --timeout-ms <N>
  • explicit overrides (still profile-adjusted)
• --no-build + --fission-bin <path>
  • skip building fission_cli and use an existing binary
• --emit-legacy-preview-artifacts
  • optional: duplicates candidate JSON under deprecated preview_* filenames (same content as nir_*).
  • default: only canonical nir_* files are written

Source semantic options

• --manifest <path>
  • defaults to benchmark/source_semantic_benchmark/manifests/smoke_windows_small_c.json
• --baseline-dir <path>
  • forwards a source semantic baseline directory to the Python runner
• --function-name <name>, --entry-id <id>, --tag <tag>
  • narrow the source semantic runner selection
• --include-ghidra-reference --ghidra-home <path>
  • opt-in reference lane only; source oracle remains the gate
• --fail-on-stop
  • exits non-zero when the source semantic go/stop decision is not go_*

---

Key outputs

Main artifacts inside each run directory:

• summary.json
  • aggregate metrics + run metadata (run_profile, target_count, timing fields)
• summary.md
  • readable summary including baseline delta and decision insights
• decision_insights.json
  • mismatch subtype ranking
  • top mismatch rows (with per-row subtype split)
  • row-level baseline/current mismatch deltas
  • go_stop_gate decision
• diagnosis.json, diagnosis.md
  • diagnosis buckets and recommended next patch
• nir_quality_candidates.json
  • per-row candidate data used for deep triage (baseline focus mode reads this path)
• Legacy (opt-in): preview_quality_candidates.json, preview_explicit_blocked_candidates.json, preview_explicit_aligned_candidate_report.json — only when --emit-legacy-preview-artifacts is set

---

Recommended iteration workflow

1) Fast loop (developer inner loop)

cargo run -p fission-automation -- nir-check \
  --lane nir \
  --run-profile fast \
  --focus-top-mismatch 5 \
  --no-build \
  --fission-bin ./target/debug/fission_cli \
  --baseline benchmark/artifacts/automation/latest/nir/summary.json

Use this to quickly validate whether the top mismatch rows move.

2) Mid loop (PR-quality check)

cargo run -p fission-automation -- nir-check \
  --lane nir \
  --run-profile mid \
  --no-build \
  --fission-bin ./target/debug/fission_cli \
  --baseline benchmark/artifacts/automation/latest/nir/summary.json

3) Full loop (periodic broader validation)

cargo run -p fission-automation -- nir-check --lane nir --run-profile full

---

Interpreting the go/stop gate

decision_insights.json includes:

• go_p5h3g_candidate
  • mismatch improved under safe dominant subtype conditions
• stop_hold_p5h3f
  • no meaningful mismatch reduction or risk signal is insufficient
• stop_no_baseline
  • baseline comparison unavailable

Treat this as an operational guardrail, not an absolute truth. Always inspect top mismatch rows and subtype distribution together.

---

Development notes

• Core modules:
  • main.rs + cli.rs: thin entry and Clap surface
  • lanes/nir_check.rs: NIR lane pipeline; lanes/mod.rs: manifest/target resolution
  • artifacts.rs: JSON/Markdown emission for a run directory
  • gates.rs: go/stop exit code and performance regression checks
  • report/: summary/delta/decision-insight construction and markdown rendering
  • diagnosis.rs: diagnosis buckets and recommended patch classification
  • corpus.rs: candidate aggregation and quality artifact assembly
  • inventory.rs: fission_cli inventory execution and loading
• Default automation config ships with this crate under crates/fission-automation/config/
  • sentinel_sets.toml: lane target definitions (paths inside are repo-root-relative)
  • preview_explicit_source_inventory.json: optional source inventory overlay (if present)
• Tests:
  • cargo test -p fission-automation

This site is open source. Improve this page.