Track and consolidate .claude, and add a pipeline for implementing FHIRPath issues - #2699
Open
piotrszul wants to merge 18 commits into
Open
Track and consolidate .claude, and add a pipeline for implementing FHIRPath issues#2699piotrszul wants to merge 18 commits into
piotrszul wants to merge 18 commits into
Conversation
These files were invisible to git because of a blanket /.claude/ rule in .git/info/exclude (a local, untracked exclude file). Checking them in as-is so we can review and decide what should actually be tracked, ignored, or removed. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
The fhirpath implementation guide and code review checklist were personal working documents, not project-shared content, so they belong in the local .local/_claude area rather than under version control. Also dropped the now-dangling reference to them from CLAUDE.md. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
The spec files in .claude/specs/ duplicated content already tracked under .claude/skills/fhirpath-spec/references/, and the remaining files each belonged conceptually to a specific skill rather than a shared top-level folder. Moved DSL_Testing_Strategy.md into fhirpath-test-designer and created a new fhir-search-spec skill for the FHIR search specification and SearchParameter registry, then updated CLAUDE.md to point at the new locations. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
The design document isn't referenced by any other file, so drop the pointers to it from CLAUDE.md and the fhir-search-spec skill. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Now that CLAUDE.md and the fhir-search-spec skill no longer reference this document, and nothing else in the repo does either, drop the file itself. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Turns a FHIRPath issue into a PR ready for review, so the 30-odd open feature issues can be worked without re-deriving the process each time. An orchestrator skill drives the run and delegates to specialists for spec lookup, test design, exclusion hygiene, and correctness review. Gates are mechanical rather than judgment calls. A change confined to a new function on an existing provider proceeds on its own; anything touching the parser, operators, evaluation context, or type system stops and escalates into an OpenSpec change for design review before code exists. Ambiguity between the spec and the reference corpus also stops. The test designer skill described a different project's DSL, so it would have emitted tests that do not compile. It is rewritten against the real one, and a contract test now exercises every documented construct to fail the build if the two drift apart. CLAUDE.md loses the process it duplicated, along with stale versions and guidance inferable from the repo, and keeps the gotchas that are not: formatting gates compilation, stale upstream modules surface as bogus compile errors, and the conformance baseline asserts how each excluded case fails, so an obsolete exclusion turns the build red by itself. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
… reference DSL_Testing_Strategy.md documented sb.extension(...) as a DSL builder method, but FhirPathModelBuilder has never had one — extensions are accessed by URL and serialised the way HAPI does, which the map-based synthetic subject cannot express. Corrected the guidance to point at withResource, matching what the skill's own SKILL.md already says about the map-based builder's limitations. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
fhirpath-spec previously relied on an untracked, per-machine symlink at .local/fhirpath.js to consult the official reference implementation. That path is invisible to worktrees and to anyone else who checks the repo out, so the cross-referencing capability silently vanished outside the original author's own checkout. cache-github-repo replaces it with a tracked, per-project pin (.claude/repo-cache.yaml, keyed by org/repo) and a version-keyed local cache shared across projects and worktrees. It separates the interactive decision (which version to use, asked once via `configure`) from the non-interactive fast path other skills delegate to (`ensure`, which fails outright rather than guessing when run unattended). fhirpath-spec's Step 0 and implement-pathling's --unattended propagation are updated accordingly. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Clean-session testing surfaced three ways these skills skipped the interactive version-pin prompt instead of actually asking: silently picking the latest tag, misreading fhirpath-spec's --unattended passthrough as unconditional, and treating a general "auto mode" disposition as equivalent to the skill's own --unattended flag. Rewords the ensure/configure branches and the reference-implementation check so each of these no longer reads as a valid reason to skip the prompt. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
fhirpath-test-designer's test-matrix review gate had no defined behaviour under --unattended, unlike the other gates in implement-pathling's pipeline, so an unattended run could stall waiting for a reviewer that will never respond. Also removes a hardcoded model name from the commit templates and trims prose that duplicated content already stated elsewhere in the same skill. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Records the fhirpath.js version pin so the cache-github-repo skill has a deterministic, shared reference across sessions. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
An adversarial review against the writing-great-skills rubric found: Step 5 and Step 6 disagreed about when tests get written, the branch-exists stop in Step 2 was missing from the Step 0 gate table (leaving --unattended behaviour undefined for it), and the Reminders section duplicated several sentences already stated in the numbered steps. Also disclosed the rarely-taken OpenSpec escalation procedure to a separate file so the common path stays legible, and pointed two build-gotcha notes at the parent CLAUDE.md instead of restating them. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Review now runs before the branch is pushed, so findings are folded into the commits they belong to and the pull request opens in the state it is meant to be judged in, rather than accumulating corrections after the fact. Adds a liveness check before branching, so an unsupervised run cannot spend itself reimplementing work that has already landed, and treats issue text from the public tracker as scope to act on rather than as instructions to follow. Moves the build ladder and the commit and pull request templates into reference files, matching the layout the other Pathling skills use. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The DSL contract test left two documented surfaces unexercised, so removing either would have kept the build green while the test-designer skill went stale — the drift the test exists to prevent. It now covers them. The skills also carried three inaccuracies that would mislead a run rather than merely read poorly: the empty-versus-absent distinction was described two incompatible ways across one skill, the unattended flag was not threaded to the reviewer subagent that consults the spec skill, and a section headed "fields that do not work" listed one that does. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Step 2 hardcoded origin/main as the branch point, which is correct for real issue work but breaks battle-testing the pipeline's own tooling before it merges to main — a branch built from origin/main would be missing the very skill files the run depends on. --base makes the override explicit and threads it through Step 10's review range and Step 11's PR target instead of relying on a manual, unrecorded workaround. Also renamed the worktree path from .worktrees/<N> to .worktrees/issue/<N> so it matches the branch name. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Aligns the implement-pathling skill's --worktree path with the EnterWorktree tool's hardcoded root, since that root cannot be changed via settings.json. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
When the CI R package cache misses, devtools and pkgdown are compiled from source, pulling in systemfonts, textshaping and ragg. These require font and image development headers that were not present on the runner, so the R dependency install failed with "there is no package called 'devtools'". Install the required libraries in the shared R setup action so every workflow that builds the R library is covered.
Backtick-quoting the operator in the roxygen tag produced a malformed NAMESPACE entry that R's namespace loader looks up literally, causing "object '`%||%`' is not exported by 'namespace:rlang'" even though the operator is genuinely exported. Quoting matches the existing convention used for sparklyr's "%>%" import in the same file.
|
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
A local
/.claude/rule in.git/info/excludehad been quietly hiding most of this repository'sClaude Code configuration from version control, so some skills were tracked and others were not.
This branch removes that rule, brings the hidden content under version control, tidies what turned
up, and then builds on it: a skill-driven pipeline that takes a FHIRPath issue to a PR.
The two halves are related — the pipeline is only shareable because the configuration it lives in
is now tracked.
Bringing
.claude/under version control.claude/specs/into the skills that actually use those references, and add afhir-search-specskill for the FHIR search specification and SearchParameter registry.A pipeline for FHIRPath issues
/implement-pathling <issue>drives an issue to a PR ready for review. It delegates to specialistsfor spec lookup, test design, exclusion hygiene, and correctness review, and stops at the PR — it
does not merge.
Gates are decided by paths rather than judgment. A change confined to a new
@FhirPathFunctionmethod on an existing provider proceeds on its own. Anything touching the parser, operators, the
evaluation context, or the type system stops and escalates into an OpenSpec change, so a
framework-level design is reviewed before code exists. Unresolved ambiguity between the
specification and the reference corpus stops too.
Correctness review runs before the branch is pushed rather than after, so a finding is folded into
the commit it belongs to and the pull request opens in the state it is meant to be judged in.
Two further gates guard the entry point, because most of the run is unsupervised. A liveness check
confirms the issue is still open, that no pull request already covers it, and that the functions it
names are not already registered — so a run cannot spend itself reimplementing work that has
landed. And since the tracker is public, issue text is read as scope rather than as instruction:
text that directs the run is reported back rather than obeyed.
Three supporting skills:
pathling-yaml-exclusionsfor the conformance baselines,pathling-fhirpath-reviewfor a correctness rubric, and a repairedfhirpath-test-designer.Repairs found along the way
fhirpath-test-designerdescribed a different project's DSL. It namedFhirPathTestBase,com.example.fhirpathand@TestFactory, and stated thatstringEmpty(),elementEmpty()andchoice()do not exist. Pathling usesFhirPathDslTestBase,au.csiro.pathling.test.dsland@FhirPathTest, and all three methods do exist. Any generated test would not have compiled.DslApiContractTestnow exercises every documented construct so the two cannot drift apartunnoticed.
spark-catalystdeclaredname: Apache Spark Catalyst API, which is not a valid skill name.CLAUDE.mdcarried stale versions (9.3.0-SNAPSHOT, server1.0.0-SNAPSHOT; actually9.9.0-SNAPSHOTand2.0.1) alongside a feature-implementation process now owned by theorchestrator. It is rewritten around gotchas that are not inferable from the repository, and
hardcoded versions are dropped rather than updated.
A note on the diff size
77k insertions, of which ~75.8k is vendored specification text that was previously untracked:
search-parameters.json(67.6k),FHIR_search.md(3.5k),FHIRPath.md(4.6k) andFHIR_FHIRpath.md(0.7k). Roughly 2k lines are authored content. Everything reads as an additionbecause these files were never tracked —
CLAUDE.mdappears as a new 86-line file even though itis a rewrite of a 438-line one.
Verification
DslApiContractTest— 10 tests, 0 failures. Every DSL construct documented in the test-designerskill compiles and behaves as documented.
YamlReferenceImplTest— 1821 tests, 0 failures, 961 skipped (the exclusion baseline).YamlFhirPathTest— 1268 tests, 0 failures.FhirViewShareableComplianceTest— 124 tests, 0 failures.namematches its directory; no dangling references.spec-drivenartifact graph isproposal → {design, specs} → tasks, not the linear order the schema listing suggests.The orchestrator itself has not yet been run end to end. The intended first run is against #2385
(existence functions), attended, before trusting it unattended.
🤖 Generated with Claude Code