Require authenticated sidecar cache purge before a deletion can complete and keep failures retryable. Mount tombstones outside restored application data while leaving deletion disabled by default.
The structured model and StructuredCvProfileJson.FromSections already map
Projects/Certifications/Languages headings, but the AI normalize prompt
never emitted them, so on the benchmark CV the entire Projects section and
the in-summary languages (English Native, Norwegian B1) were silently
dropped. This closes that gap upstream — no backend schema or data change.
ai-service (tools/summarizer/app.py):
- /cv/normalize: added # Projects and # Certifications headings; a
languages-from-prose rule (pull "native English", "Norwegian at B1" out
of the summary even with no Languages section; ignore programming
languages); and skill-group prefix stripping ("Development:",
"DevOps & Infrastructure:", "Practices:" dropped, only the skills kept).
- /cv/classify-block: Projects and Certifications added to the section
enum + rules (fallback path).
Backend:
- LooksLikeNormalizedMarkdownCv now recognises # Projects / # Certifications
so those CVs still take the markdown assembly path.
Tests:
- CvExtractionCoverageTests (4) lock the C# mapping of Projects,
Certifications and Languages sections into the structured profile.
- ai-service test_classify_block_supports_projects_section (1).
426 backend tests, 17 ai-service tests pass; app.py compiles.
The LLM behaviour (prompt -> headings) needs Ollama to observe and was not
run here; the C# side that consumes the headings is proven and the prompt
change is additive. Merge-not-replace + the review screen are the next
increment (2.1-a, approved: always-review, conservative merge).
Deployment: these prompts live in the ai-service container, which
deploy.sh does not rebuild by default -- deploy with
DEPLOY_BUILD_AI_SERVICE=true or the change won't take effect.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Unblocks the documented core workflow and closes the AI-service exposure,
without changing existing behaviour.
Job/JobApplication split (additive; see ADR-002):
- New Job entity (the opportunity) with owner-scoped query filter; nullable
JobApplication.JobId FK. Nothing reads Job yet.
- Migration AddJobEntityAndProspectStages, hand-edited to drop reconciler-owned
tables the scaffolder re-emitted; verified against the real dev DB.
Pipeline: 10 internal stages across three concerns kept separate —
PipelineStage (workflow) / PipelineGroup (UI: NotApplied/Active/Closed) /
PipelineCategory (analytics). Adds Saved/Interested/Preparing/Withdrawn;
keeps Waiting and Ghosted. Kanban shows 3 grouped columns; cards keep a stage
chip and full transitions; drag applies only safe transitions (never infers
Ghosted/Withdrawn).
DateApplied nullable + SavedAt. Cleared when leaving Applied so analytics stay
accurate; the discarded date is preserved as an AppliedDateCleared JobEvent.
AI service lockdown: no host port; private ai_internal network (backend is the
only other member); X-Ai-Service-Token required on all non-/health endpoints;
AI_SERVICE_TOKEN mandatory via compose. Verified backend-only against the live
stack.
Also carries two pre-existing working-tree files (views/ProfilePage.tsx,
views/CareerWorkspacePage.tsx) so the tree is clean for the branch integration.
Tests: +40 backend (247 total), +5 sidecar (16), +15 frontend.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Wave 4 hardening. Wrap untrusted CV/job-description/instruction text
in tools/summarizer prompts with explicit delimiters and an
ignore-embedded-instructions rule, since JD text, recruiter emails,
and free-text candidate background all flow into rewrite/normalize
prompts unescaped today.
Match score previously normalized synonyms (JS/Kubernetes/K8s/etc)
only when scanning the job posting, not when checking the CV corpus,
so a CV using an abbreviation the job spelled out never matched.
SkillTagger.MatchesTag reuses the same synonym regex for both sides.
The structured /cv/* calls funnel through a provider router so production can
offload a weak local GPU (GTX 1060) to a cloud provider without any .NET change.
Default stays "ollama" (keyless/local) and /summarize remains local distilbart.
- AI_PROVIDER=ollama|gemini|groq dispatch inside _ollama_generate_json/_text
(entry-point names kept, so no call sites change; Ollama path is byte-identical).
- Gemini (x-goog-api-key header, not URL query) and Groq (OpenAI-compatible
chat/completions) added via stdlib urllib — zero new dependencies.
- /health reports ai_provider + ai_provider_configured.
- Keys read from env only; never logged/committed.
- Compose + .env.example pass AI_PROVIDER/GEMINI_*/GROQ_* through.
Tests: 11 passed (default Ollama unchanged, Gemini/Groq dispatch, missing-key 503,
health reports provider).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
cesnimda
cesnimda