Structural / critical-path track — risk model & intent interrogation
Depth for the structural finding class (catalog category 14, OPT70–OPT75 plus
OPT78, OPT83, OPT84, OPT85 and OPT87) — the findings routed from the measured critical path rather than a YAML
match. OPT74 and OPT87 have no router yet: they emit no finding, prompt or Risk row, and their catalog entries are applied by hand. OPT85 is routed by a timestamped drill-log leaf. OPT83 is a finding dict (risk, guardrail and rollout on the record)
rendered as its own card at the pole, uncredited: it does not rank in the
findings table or get a Risk row, and its prompt carries an independence
re-check rather than the history/intent interrogation below. Its gates fail
closed: any fact the workflow file cannot prove, including a candidate step that
sets continue-on-error, holds the pole back (the full list is in the OPT83
catalog entry). OPT84 (a slow step on the pole runs a tool with a native drop-in)
is the same shape: a finding dict rendered as its own uncredited card after any
OPT83 card, its only number the step's p50 as a ceiling. The card carries no
prompt block: it links the target's docs (the oxfmt and oxlint migration
guides and overview pages; the TypeScript 7 row links the release
announcement, as there is no migration guide) with a benchmark-first
instruction and the no-weakening rail. Its gates fail closed too: a scan with
no ESLint config read or a failed config reader (the two scan gates it shares
with OPT82), an unreadable package.json or Prettier config on the Prettier
step's walk, and a TypeScript version it cannot pin down all hold the step
back; the full held-back list is in the OPT84 catalog entry. The report
verifier re-derives each tool row's fit from the config facts stamped on the
finding. OPT78 is routed by a drill-log leaf rather than the
critical-path router, so it emits no finding dict and the render boundary below
never sees it:
it gets no Risk row or banner. Its HIGH risk is stamped on the pole's drill-down
instead, and its intent check, guardrail and rollout ride in the agent prompt
blocking_path.py builds for the vitest-isolate-pool leaf.
Routing mechanics and the full risk model live in ARCHITECTURE.md §11; this
doc covers what the render boundary enforces and the intent check baked into
every per-finding prompt. A normal run does not construct any of this by hand
(the renderer bakes it into each prompt) — read it for depth.
Contents
Risk is mandatory and renders loud (but never demotes the rank)
Structural levers can degrade correctness, not just performance, so every
structural finding carries a risk (LOW/MEDIUM/HIGH), a mandatory
guardrail, and a rollout — and the render boundary rejects any
structural finding missing risk or guardrail (a future edit can't silently
turn a HIGH-risk lever into a safe-looking one). They rank in the same findings
table by Δ wall-clock — the biggest win is usually the slowest gating check, a
structural lever, so risk does NOT demote it; instead the risk is loud where
it renders (a Risk row, a 🔴 HIGH banner, and a prompt making the agent state
the failure mode + fallback + rollout). The canonical case is scoping a
build/test to "only what changed" (turbo --filter / nx affected /
vitest --changed, OPT70): the biggest wall-clock lever and the most dangerous
change — it can miss a transitive dependency, silently drop coverage, or turn a
build/import error into a false pass. NEVER present it as a safe quick win:
every such finding states its failure mode, a full-build/full-suite fallback, and
a parallel-run rollout.
Interrogate the target file's history & intent
A detector firing says a pattern matches; it does not say the matched code is there by mistake. So every per-finding prompt instructs the user's agent to read the target file's git history and recover its intent before shaping a change:
git log -p --follow -- .github/workflows/<file>.yml
gh pr list --search "<file/feature>" --state allMany waste-looking patterns are deliberate — a check on every push may keep triage labels/comments fresh; a "redundant" full build may back a correctness gate; a removable-looking step may be load-bearing for a security split (unprivileged exec + a privileged labeler). Recency matters: code shipped/iterated in the last days/weeks by an active owner is almost always intentional. Decision rule (in the prompt): if the fix contradicts the evident intent, the agent does NOT emit it as a quick win — it flags a behavioral/policy change needing the owner's sign-off (documenting the intent it changes) or returns an explicit "skip, here's why". The prompt carries this rule so the agent applies it without re-deriving it.
Verification bar
A scoped local test is pre-flight evidence, not verification. A change is verified only when the repository's own CI gate passes on the pushed branch, or when exactly the commands that gate CI pass against the change. The hand-off states which path proved the change.
The agent also states the pole's observed range beside the predicted saving. When the predicted saving is smaller than normal run-to-run spread, it does not open a pull request: the expected change is too small to distinguish from ordinary noise.