Repository navigation
feat: serve structured rule content and grade with program-calls (ADR-0024) - #625
Conversation
rules gains content and rule: with content each selected rule carries the workshop-rs rule-content-v1 document plus statement spans, bounded to 20 rules by default and always selection-wrapped. Adds wright inspect rules, schema and contract docs, and a conformance test against the owner's published schema. OverPy results report names as unavailable pending the provider capability. Bumps workshop-rs to 1.16.0; its span pruning makes the source-map-span-overflow warning unreachable for the covered input, so that test now asserts the load still serves.
…(ADR-0024) program-calls compiles the entry with the oracle or Wright, reads the Workshop text through rules with content, and counts calls by catalog id and positional argument patterns. A Wright that does not serve rule content makes the check unavailable instead of failing or passing it. Fixtures cover the Proximity-repel rule in OverPy and in en-US and zh-CN Workshop, with a one-short-stun negative.
rules.rule with a numeric id now joins cfg in the stale-id refusal after a reload; an unmatched name maps to invalid-selection like other selection fields while a duplicate name keeps the ambiguous-rule refusal cfg reports. Extends the CLI/agent selection parity test to inspect rules with content/max/rule, adds the provider test asserting unavailable.names=opy-rs over canonical content on an OverPy input, covers the 20-rule default bound and selection totals, and documents the unavailable member in the agent schema.
The ruleContent cache key now includes maxRules so two checks in one scenario with different bounds cannot share a truncated read. New tests pin that a withheld selection fails the check instead of passing on half the program, and that a Wright unable to serve rule content reports unavailable rather than a pass — whether it ignores the field or refuses the request.
…ostic change workshop-rs 1.16 prunes provenance spans that do not resolve in an attached source, so the source-map-span-overflow warning no longer enters the cli envelopes: cli:check drops 514B/128.5T/diagnostics:1 to 193B/48.2T/diagnostics:0 (the pre-#619 size), and cli:lint, cli:analyze, cli:compile record diagnostics:0. compile emittedBytes 341401->341411 and small envelope shifts ride along as in-band accepted values from the same run.
#624 requires the driver test to prove resolved span paths on every statement, not a single one; the selected rule's conditions and actions all carry declarations-rules.ws.
e54-bot
left a comment
e54-bot
left a comment
There was a problem hiding this comment.
program-calls counts disabled calls (benchmarks/agent/bench_grade.py, count_calls).
count_calls walks content["conditions"] and content["actions"] and counts every matching call without reading disabled. workshop-rs 1.16 emits disabled at three levels: the rule document (required by the schema), each condition, and each action (rules/content.rs:61,117,229). A disabled call never runs, so an agent that disables a required applyImpulse or setStatusEffect still passes the check. That defeats the structural grading in ADR-0024 decision 5 and SPEC-414 REQ-010. Neither count_calls nor the new tests mention disabled.
Correction: skip statements with disabled: true, and skip all statements of a rule whose content["disabled"] is true. Add a test with one disabled action.
Nothing else found against #624's acceptance criteria.
count_calls counted calls inside Disabled() actions, disabled conditions, and disabled rules — none of which run — so an agent could satisfy a check with dead code. Rules with content.disabled, and statements carrying the rule-content-v1 disabled marker, are now skipped; a test pins one disabled action, one disabled condition, and one disabled rule not counting.
|
Review finding addressed in 7938cc1 (pushed):
Verified: the 8 program-calls tests pass, including the repel fixtures in all three surface forms and the one-short-stun negative. Follow-up review: e54-bot authored the PR so GitHub refuses a native re-request (422); flagging readiness here instead. |
e54-bot
left a comment
e54-bot
left a comment
There was a problem hiding this comment.
LGTM. 7938cc1be skips disabled rules, conditions and actions in count_calls, and the new test covers all three levels; CI is green. (Posted as a comment because this identity is the PR author and cannot approve.)
2 participants
Closes #624.
Implements the Wright side of ADR-0024.
Changed
rulesgainscontentandrule: withcontent, each selected rule carries theworkshop-rs/rule-content-v1document plus statement spans; default bound of 20 rules, always selection-wrapped;ruleselects exactly one. OverPy results reportunavailable: {"names": "opy-rs"}.wright inspect rules --content/--rule/--only/--file/--max; agent schema (including theunavailablemember),agent-contract.md, CLI docs, and the MCP tool description updated.rules.ruleby numeric id joinscfg's stale-id refusal after a reload; an unmatched name reportsinvalid-selectionlike other selection fields; a duplicate name keeps theambiguous-rulerefusalcfggives.program-callscheck kind (catalog id + positional argument patterns),unavailableon a Wright without rule content, read content cached by (source, maxRules); Proximity-repel fixtures (OverPy, en-US, zh-CN) and a one-short-stun negative.workshop-rs1.12.0 -> 1.16.0.overwatch-ai-pvere-baselined (cli:check514B/128.5T/diagnostics:1 -> 193B/48.2T/diagnostics:0;cli:lint/cli:analyze/cli:compilediagnostics:0;emittedBytes341401 -> 341411) — span pruning removes thesource-map-span-overflowwarning from the envelopes.Verification
cargo test --workspace --all-targets --all-features,cargo clippy --workspace --all-targets --all-features -- -D warnings,cargo fmt --all -- --check,git diff --check: pass.python3 benchmarks/metrics/metrics.py checkover the full corpus: within bands after the re-baseline;cli:checkwas the only out-of-band metric.wright serveon the repel OverPy input:unavailable: {"names": "opy-rs"}with canonical content and resolved span paths.inspect rules --contentwith--maxand--rule; every statement'sspan.pathresolves.workshop-rspublishes (contract test).Follow-up
session.rsis unreachable under workshop-rs >= 1.13 span pruning:SourceMap::applywrites spans only into program provenance,set_file_sourceprunes exactly the spanscheck_span's byte-range audit rejects, and every othervalidate()failure is source-independent — the unretained re-parse can only re-surface the same error. Removal tracked in Remove the now-unreachable retained-source validation fallback in session.rs #626.