Spec Trace

Question: $ARGUMENTS

Produce an ordered chain of spec algorithm invocations from a web-visible entry point to the behaviour in question, quoting the deciding step at each hop, and end in a verdict. The output goes into a Bugzilla comment as an argument, so it has to survive a reader checking it. Pass --gecko to also map the trace to Gecko source.

webspec-index trace finds the routes and supplies each hop's step number, verbatim text, guard steps and call-site link. Your work is judging which of those routes the scenario actually takes.

Phase 0: Resolve

• Entry anchor — the API the scenario starts from. webspec-index idl 'Interface.member()' resolves IDL members directly; use search or anchors for non-IDL entry points.
• Target anchor — what the question asks about. Events are indexed as definitions: webspec-index anchors '*eventname*' --spec HTML.
• Scenario facts — the preconditions that decide branches. State them in the output. A trace is valid only for the scenario it assumed, and most wrong traces are wrong because an unstated assumption silently picked a branch. Where the question leaves a branch-deciding fact open, name the assumption and trace the branch the reporter most plausibly meant.

Phase 1: Enumerate routes

webspec-index trace '<SPEC#entry>' '<SPEC#target>' --max-depth 9 -l 40 --format markdown

Nine hops is not generous; real chains run that long. Too low a depth reports zero routes, which reads exactly like a genuine NOT REACHED, so raise it once before believing an empty result.

--kind step is the default, so prose mentions, notes and IDL tables are already excluded — around 60% of the reference graph, none of it calls.

Read the header before the routes:

• Found N trace(s) with truncated: false means the enumeration is exhaustive at that depth. This is what lets a NOT REACHED verdict claim more than "I did not find one".
• Search was truncated means it is not. Raise -l or narrow the endpoints; if it still truncates, the verdict is INCONCLUSIVE.
• indexing N spec(s) the route touches on stderr means specs a route crosses were fetched on demand, costing about a second each on a first run. If it keeps appearing across runs, the chain reaches further than the search will follow, and the verdict is INCONCLUSIVE.

When the target turns out to be reachable only from somewhere other than your assumed entry point, refs '<SPEC#target>' --direction incoming --kind step -l 50 shows who really calls it.

Phase 2: Judge each route

The tool proves a route exists in the reference graph. It cannot know your scenario. Decide route by route whether it is taken, from what each hop already carries:

• Guards. Does a - under: condition contradict a scenario fact? That kills the route.
• Arguments. Does the step pass, or omit, an optional argument the callee branches on? An omitted argument frequently disables the branch the whole question turns on.
• Ordering. Compare step numbers within a section. A call at step 20 runs before one at step 24, so state it depends on may not be set up yet, or may already be torn down.
• Identity. Does the hop act on the object the question is about? An algorithm that aborts the previous ongoing navigation does not report this one.

Read the whole section when a hop's own text and guards are not enough: webspec-index query '<SPEC#anchor>' --format markdown.

A NOT REACHED verdict needs the step where the chain stops, and that step is never in the output. trace enumerates routes that reach the target, so the step that ends a chain instead — returning, aborting, or taking the other branch — has no edge to the target and cannot appear in any route. It is invisible to the search by construction, and it is usually the step the answer turns on. Read the last section a surviving route reaches, starting at the step after the one the route left off at, and quote what you find there.

Keep the reason each route was rejected. Those reasons are half the output, not scratch work: they are what makes a NOT REACHED verdict an argument, and what a reporter will push back on.

Phase 3: Render

**Question:** <the question>

**Scenario:** <the preconditions assumed>

1) [`<SPEC#algorithm>` step N](<call-site url>) calls `<SPEC#callee>`, <argument fact that decides the branch>
   - under: <guard>
   > <verbatim step text>

2) …

**Verdict: <REACHED | NOT REACHED | INCONCLUSIVE>** — <one sentence>

| Rejected route | Rejected because |
|---|---|

trace --format markdown already emits the numbered hop list; add the scenario, the verdict and the rejected-route table around it. Say explicitly where a chain returns to an earlier algorithm — that unwind is usually where the answer lives.

Every link resolves to an element id, so exists can check it and a reader can tell when it breaks. Two forms, preferred first: the call-site id the generator emitted for that mention, which trace supplies and which lands the reader on the calling line; failing that, the section anchor with the step number in the link text — [`HTML#navigate` step 24.2](https://html.spec.whatwg.org/#navigate). About a third of hops have no call-site id, and a step that calls nothing never does, so the deciding step of a NOT REACHED verdict usually takes the second form.

Exactly one verdict:

• REACHED — the spec specifies the behaviour; name the step. Firefox not doing it is a bug.
• NOT REACHED — no route survived judgement. Consider whether the spec omission is itself the defect, in which case Firefox is compliant and the follow-up is a spec issue.
• INCONCLUSIVE — the chain hit a UA-defined step, an unindexed spec, or a truncated search. Say where it stopped and what would resolve it. Never soften this into another verdict for tidiness.

--gecko mode

Append a table mapping the trace to Gecko source, keyed on canonical URL:

searchfox-cli --spec-refs '<url>'

• Separate code hits from test hits, and call out algorithms with no hits — an unreferenced spec algorithm is often the reason for the bug.
• One row per distinct URL, not per hop. An algorithm the chain re-enters collapses into one row listing the hops it covers.
• Add rejected algorithms as (not reached) rows. For a NOT REACHED verdict they carry the weight: the code exists and simply is not wired to this route.
• Discard matches under .claude/skills/ and .agents/skills/. In-tree skill documentation quotes spec URLs in its examples and is indexed as code, in both copies.

Hard rules

Every quote comes from tool output. trace and refs emit plain step text; use it as given. query renders markdown, so its step text carries *emphasis* around variable names and inline [links](...); stripping that to recover the spec's prose is not paraphrasing, changing a word is. A truncated fragment presented as a quotation is an invented quote, and an invented quote loses the argument the trace was written to win.

Validate every anchor with webspec-index exists '<SPEC#anchor>' before it ships, the rejected-route table included. Exit code 0 means it resolves.

Trace the spec as of today. If the bug references an unlanded spec PR, add --pr N to query, refs and exists, and say in the output which PR the trace assumes. trace has no --pr; against a PR, walk refs --kind step in both directions instead.