Skip to content

Implementation plans

Plans live in docs/src/implementation-plans/. 306 total.

Delivered (237)

Title Issue Last updated Summary
Travel traveller resolution: -1 / null sentinels — Implementation Plan #1153 2026-08-14 Task-by-task plan replacing other_internal/other_external with '-1' and JSON null across the ingestion provider, create/response DTOs, the repo-level resolution comment, an Alembic data migration, and the already-shipped frontend selector/resolver/table/chart consumers.
Travel traveller resolution: replace string sentinels with -1 / null — PRD #1153 2026-08-14 Replace the other_internal / other_external user_institutional_id hack (issue #1153) with a deterministic, backward-compatible representation: '-1' for explicit Internal other, JSON null for explicit External other, any other value is a source SCIPER resolved against Headcount scoped to (carbon_report_module_id, sciper).
Train CSV station disambiguation — required country_code #1183 2026-05-29 Sources the train location seed from the trainline-eu stations.csv (ISO-2 country per station) and makes country_code REQUIRED in the train CSV trip resolver — no more CH default. Retires the CH-backfill script.
1186 — Train natural_key validation: close the remaining gap #1186 2026-08-20 Closes the remaining natural_key validation gap #1183 left open: train API creates and CSV not_found stations no longer persist silently with zero emissions — both now fail hard at ingest time.
Restrict backoffice year creation to [2025, current_year] #1204 2026-07-07 Cap the backoffice year selector and year-creation endpoint to years from 2025 through the current year, removing the leftover 2023/2024 test range.
Move 'Incomplete' tag computation to backend #1215 2026-05-22 Backend year-configuration response emits explicit incomplete + reasons per submodule/module. Frontend deletes its client-side derivation. Only factor + reference uploads are mandatory.
BackOffice upload card — API success status + FACTORS-first order #1216 2026-05-28 Data card now turns green when the API ingestion path succeeds (not only CSV). FACTORS-first order across Equipments / Purchases is pinned by a regression test.
1219 Stuck ingestion/recalc jobs + premature pipeline 'success' #1219 2026-05-19 Stage incident: a zombie emission_recalc row tripped uq_emission_recalc_active, the IntegrityError poisoned the runner's job_session, and the job never reached FINISHED — self-propagating the stall. Fixes runner session resilience, adds dedup to the csv/api fan-out, and makes pipeline completion server-authoritative with a 3-phase UI.
1234 — Pipeline operations console #1234 2026-11-18
1236 — First-class pipelines table #1236 2026-05-20
Lighthouse: single config + nightly cron #1332 2026-06-08 Merge the duplicate Lighthouse configs into one source of truth and move CI from per-PR to a nightly cron.
Clarify re-upload vs recalculate messaging on partial CSV ingestion #1398 2026-07-07 Partial/failed CSV ingestion jobs must tell the user to re-upload the file, not just recalculate, when rows were skipped or never reached.
Split Grafana p99 Latency Alerting by Endpoint Class; Add GlitchTip Alerting #1402 2026-08-24 Separate p99 latency alert thresholds for upload/job/pipeline endpoints from normal API calls, exclude SSE streams from duration-based alerting, and add error-rate alerting on GlitchTip.
1411 — Hide module-status validate button from standard users + consolidate frontend permissions #1411 2026-06-09
CSV Provider Consolidation Implementation Plan #1458 2026-06-10
Buildings room-add disclaimer fires inconsistently #1463 2026-07-07 The Archibus disclaimer toast is coupled to the entire postItem() promise, including unrelated downstream refresh calls, so any transient failure among them silently suppresses the disclaimer even though the room was created.
Factor/entry join-key normalization #1489 2026-09-08 Shared pydantic field types normalize every factor-resolution join key symmetrically on both DTO families (strip, lowered currency/cabin class, uppered country codes with the RoW sentinel kept, spreadsheet numeric ids coerced, blank optional keys to None). Pre-existing rows are audited by an operator script instead of rewritten by a migration: on prod and stage nothing a lookup uses differed.
Fix silent recalculate failures and wrong step count in the recalculate progress UI #1523 2026-07-07 Surface SSE connection errors during 'Recalculate' as a visible notification, and give recalc-triggered pipelines their own single-step phase label instead of borrowing the 3-phase ingest pipeline's 'Step 1/3 - Inserting data...' text.
Reference-data CSV upload: fail on unexpected column names #1545 2026-08-20 ReferenceDataCSVProvider silently logged (never raised) on unrecognized CSV columns, so a misspelled column resolved to None on every row with no error. Made unknown columns a hard failure; confirmed reduction_objective CSVs already validate column names correctly (every field is required).
API Connect: connectors and encrypted connections — implementation plan #1552 2026-07-06 Tiered TDD plan for a modular connector registry, one encrypted connection per connector (entered in the form), per-module connector_luid datasources, the travel + new headcount providers driven from them, and the shared Tableau logic in one base provider.
API Connect: connectors, connections and per-module datasources #1552 2026-07-06 Make the backoffice 'API connect' form functional via a modular connector registry. Store one connection per connector (server/site/username + connected-app credentials entered in the form, secret encrypted at rest), bind a datasource LUID per module, and drive the travel + new headcount providers from one shared base provider.
Simulator Plan — Backend #1556 2026-07-16 Backend architecture for the Simulator Plan module: plan-year reports as ordinary CarbonReports, reference-year factor resolution, per-module planner schemas, and the API surface for the year-range lifecycle.
Simulator Plan — frontend follow-ups #1557 2026-07-23 Four planner-frontend pieces: the type-2 prefilled slider table (delivered; ModuleTable decomposition deferred to its own plan), the wrong-plan report resolution fix (delivered; overlapping-plans regression test outstanding), dropdowns resolving factors from the plan year instead of the reference year (delivered), and the Comment button disabled by the Calculator's validated state (delivered).
Issue 1558 — Calculator year selector doesn't show a newly-opened year until a hard refresh #1558 2026-07-08 yearConfigStore.configuredYears (which years are open to users) was hydrated once at auth bootstrap and only ever refetched when still empty, so a year opened via the backoffice mid-session stayed invisible in the calculator's year dropdown — and could keep a user bounced to /unauthorized — until a hard reload rebuilt Pinia state from scratch. workspaceStore.units/getUnitResults were both ruled out as unrelated to this specific bug.
Idempotent tmp->processing (and processing->processed) file moves on job retry #1559 2026-08-21 Skip the move if the destination already exists so a retried ingestion job doesn't fail on a file a prior attempt already consumed.
Headcount: Allow the Same Member With Multiple Roles in One Unit #1564 2026-07-08 Headcount uniqueness is keyed on user_institutional_id alone, so a member with two SIUS roles in the same unit gets their second CSV row silently dropped; fixing that exposes a real duplicate-row risk in the Professional Travel <-> Headcount join that must be fixed in the same PR.
Stage incident follow-ups: cross-source bulk replace + stuck-job recovery #1564 2026-07-17 Headcount CSV upload after an API sync mass-skipped 8465 rows as DUPLICATE_INSTITUTIONAL_ID (bulk sources didn't replace each other), and a pod SIGTERM mid-recalc left the aggregation invisible for the 60-min stale window. Delivers cross-source full replace, 5-min heartbeat-based stale recovery, queue-time visibility, and a manual Recover button.
Fix: failed CSV re-upload blanks previous upload summary in backoffice config #1578 2026-07-07 A failing re-upload is marked as the module's current job, so the shared UploadCard hides the prior successful job's filename/rows/download info even though the underlying data is untouched — UNLESS the failed re-upload's CSV passes header validation but has every row invalid, in which case the prior data is actually deleted (see 'Caveat found during implementation' below).
Surface per-row errors from research-facilities factor recompute #1591 2026-07-07 Per-row error detail is already captured server-side when recomputing common research-facilities factors; the job just never exposes it past a bare error count, so the fix is API + frontend surfacing, not new capture logic.
Building rooms: emit a single heating leaf (electric OR thermal) per room #1575 2026-07-05 Refactor the PR #1661 heating split so a room emits only the heating leaf matching its factor's energy_type, ending the thermal/electric double-count (#1575).
Remove primary_factor_id from DataEntry.data — implementation plan #1661 2026-07-06 Task-by-task execution plan for the approved spec 1661-remove-primary-factor-id.md: bottom-up (repo → services → workflows → routes), checkbox-tracked, resumable across sessions.
Remove primary_factor_id from DataEntry.data #1661 2026-07-06 Stop persisting the resolved factor id in the entry JSON; resolve factors on demand through a per-request/per-slice FactorResolver, then flip factor CSV reupload to replace-semantics so stale factor rows are deleted instead of flagged.
SQL factor resolution for list sort/filter/pagination #1661 2026-07-06 Replace the emission-FK factor join in get_submodule_data with a classification LATERAL join so factor-backed sort/filter/search work for every row (including not-yet-computed entries); delete the Python row-loop fallback.
Rename SECRET_KEY to JWT_HMAC_KEY, split out SESSION_HMAC_KEY, prune dead env vars #1704 2026-08-18 SECRET_KEY renamed to JWT_HMAC_KEY (JWT signing only); Starlette SessionMiddleware's reused key split into a new dedicated SESSION_HMAC_KEY so rotating one signing domain never invalidates the other. Removed dead OPA_URL/OPA_TIMEOUT/OPA_ENABLED (never read — ADR-005 rejected real OPA). backend/.env.example now round-trips 1:1 with config.py in both directions.
Backend provider configuration naming cleanup #1713 2026-07-07 Split the overloaded PROVIDER_PLUGIN/ACCRED_API_URL env contract into explicit ROLE_PROVIDER_TYPE / UNIT_PROVIDER_TYPE / ACCRED_* config, renaming DefaultRoleProvider/DefaultUnitProvider and decoupling provider-selection enums from the persisted UserProvider metadata enum.
Bound background-job concurrency and size the DB pool #1723 2026-07-07 Cap per-pod running jobs with an asyncio.Semaphore acquired before the DB claim, and give the async engine explicit pool settings, to stop QueuePool exhaustion when a factor CSV upload fans out one emission_recalc child per (module, det).
Gate the per-job SSE stream's cross-unit jobs behind backoffice.configuration #1764 2026-09-02 job_stream_by_id serves two audiences (unit-facing module pages, backoffice config page) through one gate; add the one missing branch inline, no new function, no frontend change.
Project Planner: default reference year + no locked modules #1922 2026-08-04 New plan-year sections default their reference year to the workspace year the planner was opened from (sent with the range PATCH, normal prefill runs); the reference year is removable; module drawers are never locked, so a year whose reference is unset is fully usable for manual input with factors falling back to the unit's latest Calculator report year. No schema change.
Centralized purchases: manual input #1935 2026-08-19 Manual entry for the centralized-purchases submodule. Unit and coef_to_kg are entry-level fields (moved off the factor in #2078); the factor carries only ef_kg_co2eq_per_kg, matched by name. Emission = annual_consumption × coef_to_kg × ef_kg_co2eq_per_kg. Follow-up locked manual input to kg: unit is read-only 'kg', coef_to_kg is forced to 1 and hidden from the form (CSV entries keep their real unit/coef); coefficient label corrected to 'Conversion coefficient to kg'.
Sim Explorer 500/503 — READ audit on a GET #1958 2026-08-18 The 500 was a GET writing a version-chained audit row keyed on carbon_report_module_id: two parallel submodule fetches of one module collided on audit_document_one_current_idx. Fixed by deleting the READ audit and scoping audit off Simulator entries; both locks (advisory and savepoint-retry) removed as unnecessary.
Simulator Plan — Grant proposal mode (first increment) #1976 2026-08-26 Plans gain a persisted Grant proposal checkbox and, when checked, a Project Grant section (a dedicated grant carbon report) rendered before the year sections. Grant tables carry a kgCO₂eq-over-project-years column (#1979), the section carries a total grant budget with currency plus per-submodule budgets and a distribution check (#1978), Research Facilities render a platform-selection grid (#1980), Equipment gets a per-line vs global-percentage planning toggle (#1981), and results show a grant vs year-by-year comparison chart plus a grant page in the PDF (#1977).
User input for the Research Facilities module (modularity for other institutions) #2007 2026-08-24 Give research_facilities and animal_facilities a manual-entry form driven by the year's factor catalog, switch it off by default via the existing per-submodule inputs_deactivated flag, unlock that flag in the backoffice, and make it fail closed server-side.
Process Emissions: rename quantity -> quantity_kg (doc-driven schema change) #2025 2026-08-18 The data manager renamed the process-emissions quantity column to quantity_kg in the data-description doc and the SharePoint data/test/template files (issue #2025, 2026-08-17); the backend DTO, the served frontend template, and stored entry JSON still say quantity, so the current backoffice data upload fails every row. Rename end-to-end with a guarded jsonb data migration — no aliases, no dual-path.
Planner: plan rights as a policy and a permission key, not a payload flag #2043 2026-08-20 Removes the per-row can_manage flag from SimulatorPlanRead (#1930). Roles now emit a planner.plans permission key; the backend resolves one PlanPolicy per request (replacing plan_is_visible_to / plan_can_manage / require_plan_access); the frontend gates the delete button on the key plus created_by. No schema change, no product-rule change.
Background DB health poller — zero I/O on /healthz and /ready #2049 2026-09-25 Continuation of 2050's Track A1: a background loop (1 s; 30 s with a 5 s boot retry since 2026-09-25) caches a SELECT 1 verdict in a per-pod global; /healthz and /ready read it instead of doing their own DB round trip, so a saturated pool can no longer make either endpoint itself hang.
Backend compute performance — pod stability, worker split, request-path profiling #2050 2026-08-19 Six-track plan against dev-platform slowness and intermittent 504s: bound /ready and move job dispatch off API pods (Tracks A/B, PR #2081); profile compute cost (Track C), which rules out a language rewrite; then fix the simulator-plan reference-year PATCH end to end — recalc and prefill N+1s, Pydantic's per-instance default_factory tax across every SQLModel table, Core INSERT…RETURNING, and Tracks D/E's redundant recomputes — taking it 960ms → 271.4ms (PRs #2083, #2152). Track F reopens the per-year prefill fan-out behind a 21.89s dev trace whose bottleneck is not traced SQL. Track G (2026-08-18) dates ~30 real dev/stage traces against the fix-merge timestamps: every slow stage trace predates #2050 entirely (stage is still pre-#2050, unrelated to any bug); dev's two biggest traces (80s, 21.89s) predate Track F's async-job fix; but plain module-detail GETs — never in this plan's scope — still cost 1-2.2s post-fix from the same connection-checkout + sequential-small-query pattern Tracks A/C/D/E already diagnosed and fixed elsewhere, and one 32s outlier has a genuinely untraced non-DB gap needing its own repro. Track H root-causes one of those G2 traces exactly: planner_headcount was missing from get_submodule_data's is_headcount_entry tuple, so it fell through to an unfiltered whole-table data_entry_emissions aggregation (825ms) instead of the already-built, already-populated rollup-row fast path — a one-line fix, the highest-confidence item in the plan. A second 'critical' trace turned out to be stage's SQL instrumentation being off by design (C1's OTel-tax follow-up), not a new blind spot. Track I (2026-08-19) answers a process-pool/worker-count proposal against this plan's own measurements (rejected, same ceiling that closed the Rust question), confirms Track B is live on dev/stage but entirely absent on prod, and ships fixes for 310-e item 8's never-done sync-in-async audit — a synchronous Loki log handler on the root logger, an unthreaded audit→Elasticsearch sync (independently corroborated live by Track G3's own trace), and unthreaded connector-credential decryption — plus a read-only confirmation that DB-pool exhaustion is pod-local (SQLAlchemy QueuePool, 30/pod) rather than Postgres-side (27/100 in use), and a proposal to instrument the pool metric that answers that question without a manual kubectl exec next time. Track J (2026-08-19) delivers what Track H exposed and what H8 scoped: prepare_create had six silent fallbacks that published a wrong-but-plausible total (a dropped emission leaf, an unresolvable year used anyway, an unmapped type, a wrong-year factor, a factor_id resolving to nothing, an unflushed id) — all now raise, with two impossible-state guards deleted outright; and the headcount module GET's three sequential FTE round trips collapse into one GROUP BY. Lever 1 (the 289-620ms per-request connect span) deliberately ships nothing: the span turns out to wrap pool checkout, not physical connect, so G2's reading of it was wrong, and A3's 1-core recommendation is unapplyable at 3 replicas under a documented 2-core namespace quota — both remaining options need numbers that arrive free from H6's now-merged 100%-sampled SQL tracing. Track J then took the interactive write path from 50 SQL statements to 12 (24 factor lookups to 1) across eight measured steps, each holding its gain with a ratchet in a Postgres-backed statement-budget test, without making anything the caller reads back eventually consistent. J8 then found the submodule GET aggregating the entire data_entry_emissions table on every request - 444.6ms of a 475ms trace, for a submodule holding one entry - and restricted it to the module, 7.5x faster at 1M rows; that same trace also shows connection checkout at 1.7ms rather than G2's 289-620ms, which if it holds retires lever 1 entirely.
Write-path statement budget — 29 SQL statements to 12 #2050 2026-08-29 Implementation plan for the interactive write path's statement-count reduction (#2050 Track J4/I5). One headcount-member POST costs 29 SQL statements after the B3 subtree fix (50 before), and twelve of them re-read three rows because four services are each constructed with only a session and re-derive identity independently. Eight tasks, each lowering the STATEMENT_BUDGET ratchet in test_headcount_post_statement_budget_pg.py: drop two redundant session.refresh calls, skip the audit head lookup on CREATE, skip the pre-delete SELECT on create, merge the count and FTE aggregates, thread the resolved (report, project, module) through the workflow, batch factor prefetch across emission roots, dispatch the report rollup, and replace the member uniqueness pre-check with a unique index. Lands at 8, against an irreducible synchronous floor of 7 for 'insert an entry and return fresh module stats'.
Planner External AI FTE count must start empty #2061 2026-09-16 In the Planner (Grant Proposal), External AI's 'Number of users (FTE)' was pre-filled with the unit's validated Headcount total leaked from the Calculator's timeline store; the Calculator-only default is now gated off in the Planner as well as the Explorer.
Planner and Explorer results: carbon footprint per FTE chart #2071 2026-09-16 The Results page shows the unit's stacked per-FTE bar beside the main category chart. The Explorer and Planner results cards now show the same pair in the same 2:1 layout, fed by the per-FTE values the backend already derives for simulator reports.
Fix: Enter-key submit leaves a false 'Required' error on the field that was just filled #2072 2026-08-18 Submitting a module form with the Enter key adds the entry correctly but then flags the focused field as 'Required'. Quasar emits QField's blur from a 0 ms timer and QBtn[type=submit] focuses itself on Enter, so ModuleForm's @blur validation runs after reset() has already cleared the form and validates an empty field the user did fill. Gate blur-validation on fields the user has interacted with since the last reset.
Emission type resolution fails hard, and is documented #2091 2026-08-19 Every module invented its own answer for a factor CSV value the taxonomy did not know: two resolvers raised, four returned None, and four walked up the tree onto an intermediate node — which double-counts, because an intermediate node already sums its children. Measured against the real INPUT_DATA CSVs, 877 of 18,977 factor rows resolved onto a node with children; 21 of those were genuine degradations (20 headcount, 1 centralized purchase) and 16 external-AI rows collapsed into provider_others, leaving the AI breakdown chart as a single bucket. This plan makes all eight runtime resolvers raise EmissionTypeResolutionError naming the offending value, adds a funnel guard rejecting any runtime resolution onto a node with children, escalates the factor-CSV provider from skip-row to abort-upload for emission-type failures only, adds the eleven missing taxonomy leaves the real CSVs need, and ships a reference page with mermaid diagrams plus a pre-upload audit script. All 15 shipped factor CSVs resolve clean afterwards.
2096 — Empty validated modules shown as "not validated" in Results #2096 2026-08-19 Validated modules with no data were dropped from validated_buckets and the results summary, so the Results chart greyed them out with a "Validate X to see results" tooltip; fixed at recompute time, pending the admin stats-recompute backfill for reports persisted before the fix.
Headcount: user_institutional_id Mandatory Again #2138 2026-08-17 Reverts #951's decision to make HeadCountCreate.user_institutional_id optional — success criteria changed, and it must be mandatory again at creation, backend and frontend.
Issue 2146 — Lock (sub)module activation once the year is open #2146 2026-08-25 Freeze the module/submodule activation toggles in the back-office Configuration page once a year is open to users (is_started), per the 24-08 decision (Option 2). Frontend-only enforcement; allowing mid-year (de)activation again (Option 1) is deferred to post-delivery.
Manual perf-test CSV generator: one file per data_entry_type (#2161) #2161 2026-08-20 Extend backend/scripts/generate_perf_test_csvs.py from 3 coarse categories (travel/buildings/purchase) to one generator per DataEntryTypeEnum, sized at #2161's real per-type ceilings, plus a scripts/README.md covering every script in the folder.
Implementation Plan: Auto-generate TypeScript types from FastAPI OpenAPI #217 2026-05-28 Stand up a POC that pipes the FastAPI OpenAPI 3.1 schema into a generated openapi.d.ts consumed by the existing ky HTTP client. Recommends openapi-typescript over heavier SDK generators, ships the generator script, makefile target, committed snapshot, and a single migrated interface (User) as proof.
2170 — Equipment response power fields: int → float #2170 2026-08-18 EquipmentHandlerResponse.active_power_w/standby_power_w widened from int to float, matching the equipment_factors.csv contract and fixing a 500 on fractional power values.
Equipment load error: double-space kind breaks factor-map matching #2174 2026-08-19 normalize_kind() lowercased+stripped but never collapsed internal whitespace, and load_factors_map() built factor-side keys without even calling it — a double space in equipment_data.csv's equipment_class (vs. equipment_factors.csv's single space) silently failed the backoffice MODULE_PER_YEAR upload's kind→data_entry_type lookup. Also fixes the process_emissions gas-category resolver found while triaging the same make bootstrap-years log (root cause unrelated to #2174, bundled here on request).
2176 — (planner) Error 403 on usage update #2176 2026-08-19 Planner "% of reference year" slider PATCHes 403'd under the
Research facilities 500 on a closed unit's factor (#2183) #2183 2026-08-20 A research-facilities create/update 500'd on dev with a bare 'Failed to create data entry'. Root cause: #2050 Track J1's fail-hard raise for a formula that can't produce a value, hitting a factor (CLIMACT-GE, factor_id=37756) whose kg_co2eq_sum was never backfilled because its unit is closed and has no CarbonReport for the year. Three fixes: surface the ValueError as 422 with its real message, name the facility when kg_co2eq_sum is the specific gap, and make the computed backfill write an explicit 0.0 for a closed unit instead of raising and leaving the factor permanently incomplete.
Stale-chunk alert noise + blocking reload prompt #2194 2026-08-28 Every release paged the team over stale-chunk load errors (expected post-deploy SPA noise, already diagnosed in #2194) because only one of four error-capture sites in boot/sentry.ts special-cased them. Extends the existing ignoreErrors filter so they never reach GlitchTip, and upgrades the dismissible reload toast to a persistent, blurred-backdrop Dialog with Reload/Later actions.
CSV Upload Feature Implementation Summary #220 2026-04-02 This document summarizes the implementation of the CSV upload feature verification and test suite for issue #220 (DB Upload CSV - Improve Data Modifications).
PRD — CSV Upload Feature #220 2026-04-01 The CSV Upload feature allows users to load machine-generated data (emission factors, synthetic data, and other module inputs) into the database from any module table. The backend implementation, UI, permissions,...
Consolidate Alembic Files Implementation Plan #2211 2026-08-20 Collapse all 15 Alembic migrations into one autogenerated initial migration, since pre-v1.x the DB is dropped and recreated between prod/stage/dev deploys.
Centralize date input handling — format hint and validation on inline table edits #2214 2026-08-28 Editing a travel date inline in the module table gave no hint of the expected format and an unclear error on a bad entry. Centralizes the mask/validation logic used by the add-entry form into frontend/src/utils/date.ts and applies it to inline table editing too, which previously had none.
Prevent double delete in module tables #2217 2026-08-27 Clicking the delete-confirmation button twice (or pressing Enter twice) before the dialog closed fired two DELETE requests; the second 404'd because the entry was already gone. An in-flight guard in ModuleTable's onConfirmDelete plus a loading state on the submit button makes the second click impossible.
Fix headcount select going stale after a form reset #2219 2026-08-28 Standard users could only create one train/plane entry before needing to refresh the page. The shared headcount-member select in the travel-like dynamic form was keyed only on headcountMemberCount, so form.reset() after a successful submit didn't remount it, leaving it in stale state that blocked the next submission.
Faker seed-data: ~800 data_entry rows, ERD-aligned #222 2026-05-28 Right-size the random generator to ~800 rows, fix payload drift against current Pydantic schemas, wire it into the Makefile.
S3 HeadObject 404s: root cause, and the S3-vs-stop-vs-PVC decision #2220 2026-08-25 Root cause found live on 2026-08-25: the genuine 'Failed to move file' failures happen when a laptop running make dev with .env pointed at the shared dev DB polls and claims dev's ingestion jobs, then resolves their uploaded files against its own LocalFilesStore — the file sits untouched in S3 the whole time. Fixed with a fail-closed boot guard (assert_poller_isolation). The recurring HeadObject-404 error spans are expected, already-handled existence checks from the #1559 idempotent-move fix, mis-flagged by OTel's botocore auto-instrumentation — confirmed by correlating a live 'error' trace with its succeeding job. PVC is not adopted: dev already runs 3 backend replicas + 1 worker pod needing ReadWriteMany, and no RWX storage class is evidenced anywhere in the ops repo.
Explorer: show placeholders for missing additional categories #2236 2026-08-28 Toggling 'additional categories' on in the Explorer chart hid columns for categories absent from the backend response instead of showing them greyed out. Mirrors the existing main-category placeholder logic to fill in missing additional (headcount/buildings) categories too.
External AI — product-name categories replace vendor slugs #2252 2026-08-26 The AI provider categories changed from vendor slugs (google, mistral_ai, anthropic, openai, cohere, others) to the product-name spellings the re-uploaded CSVs carry: Gemini (Google), Mistral AI, Claude (Anthropic), ChatGPT (OpenAI), Copilot (Microsoft), Copilot (GitHub), Other. _AI_USE_MAP now declares only the canonical tokens of those seven spellings; the legacy slug keys are deleted (the planned DB drop removes the data that used them). provider_cohere is removed from the taxonomy, registry, and IT breakdown — Cohere is no longer a category, and Copilot (Microsoft) resolves to the provider_microsoft leaf #2091 already added, not to a repurposed cohere slot. Results-chart subcategory labels are renamed to match, and the random dev seeders emit the new spellings.
Buildings data CSV: reject rows whose room is missing from the reference #2253 2026-08-28 building_rooms_data.csv rows with a room_name absent from the BuildingRoom reference used to persist and silently compute with surface 0/None; ingestion now rejects those rows via the enrich_csv_row hook (same mechanism as the train #1186 natural-key check), so the upload finishes WARNING with per-row errors instead of a wrong total.
Headcount: members without a (known) SIUS code become Other staff (-1) #2254 2026-08-25 The Tableau sync now delivers members without a SIUS code. Ingestion (API + CSV + create/update handler) normalizes empty/missing/unknown codes to the stored sentinel -1; -1 is a display-only 'Other staff/Autre personnel' function in the table and FTE chart, not offered in the manual dropdown. Existing DB rows are left untouched.
Cache the factors query behind taxonomy lookups #2258 2026-08-26 Process-local TTL cache for ModuleHandlerService.get_taxonomy, keyed on (data_entry_type, year), invalidated on every FactorRepository write and backstopped by a 60s TTL for the cross-process case ingestion runs in a separate worker deployment. Follow-up: active cross-pod invalidation broadcast on top of the TTL, dropping worst-case staleness from ~120s to the time one broadcast round-trip takes.
Upload path fixes: auth ordering, size limit, tracing (#2261) #2261 Close auth-after-body-ingestion (a DoS surface) on both upload routes — POST /v1/files/temp-upload and POST /v1/year-configuration/{year}/upload — with an AuthFirstRoute custom APIRoute that authenticates and bounds Content-Length before FastAPI reads the body, letting the endpoints keep ordinary File/Form signatures. Adds the per-file size cap the year-configuration route never had. The double PutObject/HeadObject write lives in the vendored enacit4r-files package and is filed upstream, not patched here.
Implementation Plan: Refactor Pagination & Lazy Loading for Equipment Module #228 2025-12-16 Implement lazy loading with per-submodule pagination and sorting. Load only summary totals on initial page load, then fetch individual submodule data when users expand collapsible sections.
Explorer sandboxes are private per user #2293 2026-08-24 Two users of the same unit shared one Simulator Explore report, so each saw the other's temp entries. Explore projects are now scoped per (unit, created_by) like Planner plans: the lookup/create path stamps and filters carbon_projects.created_by, and the one-per-unit partial unique index is split per type.
CPU per request, per endpoint (#2295) #2295 2026-09-25 Local, in-process harness that measures CPU ms per request (mean, median, p95), wall ms, SQL statements, pool checkouts and OTel spans per endpoint of the ExplorerReadUser mix, at three tracing levels driven by the pods' own OTEL_* env vars, with an optional cProfile pass that splits CPU by component (SQLAlchemy, psycopg, pydantic, Starlette/FastAPI, OpenTelemetry, app code).
Load tests with locust (#2295) #2295 2026-09-24 Locust load-test suite: staged concurrency (50→1000 readers, 10→40 plan/explore creators, 5→20 parallel CSV uploads) against a COPY-seeded backdrop of N units × N years at #2161 ceiling density, with make targets for seeding, sweeping, reporting p95>1s, and dumping/restoring the seeded DB.
Perf ladder + CPU per request (#2295) #2295 2026-09-25 make perf-ladder runs a warm-up plus a stage ladder for one named config against dev, then make perf-summary computes backend CPU milliseconds per request per stage: cAdvisor CPU-seconds over the steady window divided by locust's own request count, never an OTel counter.
Escape @ characters in i18n translation strings #2306 2026-08-28 vue-i18n treats a raw @ as linked-message syntax, which compiles fine in dev but throws at render time in production; some translations also had it wrapped in smart-quoted {'@'} from a word processor. Escapes every @ as straight-quoted {'@'} and adds a regression test scanning all i18n files for unescaped @.
Logout button on the 403 page #2308 2026-08-26 Always show a Logout button on /unauthorized: a broken session (e.g. a stale account after the 1.4.1 DB reset) 403s every call, the global 403 hook redirects here without a reason, and the page previously offered only a Home button that looped straight back.
2316 — Simulator PDF export missing additional categories #2316 2026-08-27 Simulator Exploration PDF export now shows additional estimated categories in the unit chart, matching the on-screen graph and the headline total.
Headcount FTE fields capped at one decimal place #2318 2026-08-28 FTE fields in the Headcount module accepted up to 16 decimal places. Adds a generic maxDecimals config option to module fields, set to 1 on both Headcount FTE fields, enforced on the add-entry form and on inline table editing.
Centralized purchases graph grouped by product name #2320 2026-08-28 The centralized-purchases bar chart showed every bar labeled 'Unknown' because the module's grouping fell back to an institutional/factor code these entries don't carry. Adds a per-data-entry-type override so purchases_centralized rows group by their free-text name field instead.
2342 — Extract portable engineering rules into it4r-agent-kit #2342 2026-08-25 Move the portable half of guardrails.md into the shared it4r-agent-kit repo, vendored here as a plain file (not a submodule), leaving guardrails.md as the co2-specific delta.
Fix misleading default values in Explorer/Planner forms #2347 2026-08-28 Three Explorer/Planner form fields defaulted to misleading values (FTE forced to 0, no traveler pre-selected sensibly, Equipment ID blank) instead of leaving them empty or picking a sensible default. Part of the #2061 default-values cleanup.
Explore-page request stampede: in-flight dedup + mount audit #2360 2026-08-26 Dedup concurrent resolveCarbonReportId and getHeadcountMembers calls with an in-flight promise cache (11 identical report lookups and 4 identical member-roster GETs observed per explore-page load), plus a static audit of every request the simulation/explore/new mount fires and the batching opportunities it exposes.
2369 — Superadmin redirected away from non-member units (guard authorized client-side) #2369 2026-08-26 Let the backend decide workspace-unit access in the router guard: when the route's unit is not in the membership list, fetch it via GET /units/{id} (policy-authorized) instead of redirecting client-side. Fixes superadmin being bounced from units outside their membership list.
Job runner: one OTel span per job execution #2371 2026-08-26 Wrap each claimed job execution and each pipeline-reconciler sweep in an OTel span so background-job SQL nests under a parent trace instead of flooding Tempo as parentless root spans.
Propagate W3C traceparent from frontend API calls #2372 2026-08-26 Stamp a traceparent header on every ky /api request so backend OTel spans join the browser's per-navigation trace id, making GlitchTip trace_ids searchable in Tempo.
Real read policy for GET /units/{id}, and stop truncating the membership list #2379 2026-08-31 Route the unit read check through require_unit_access — the same enforcer as the workspace boundary — instead of the legacy allow-all policy stub, and drop the limit=100 pagination that silently truncated the membership list in four call sites including the session bootstrap.
Unify factor/taxonomy option delivery: one per-year endpoint + ETag, purchase typeahead, strip coefficients #2391 2026-08-31 Consolidates the seven mechanisms that feed form selects from the factors table onto the #2280 batch taxonomy endpoint (labels only, long-TTL cache + ETag), moves purchase to a server-side typeahead, and deletes the dead routes/constants the audit found. Delivered in slices; this doc tracks all seven decisions across sibling PRs.
Suppress per-chunk ASGI receive/send spans #2397 2026-08-27 One upload produces 805 spans, one per body chunk, each created and exported on the event loop. The pinned opentelemetry-instrumentation-asgi 0.65b0 supports exclude_spans natively, but only as a keyword argument — and the app is instrumented inside the FastAPI() constructor by opentelemetry-instrument, before our code runs. Shipped as an explicit uninstrument/re-instrument at boot (#2409), sequenced after #2263 so the win can be measured rather than assumed.
Comprehensive Code Review: Audit Trail Implementation (feat-240) #240 2026-02-16 The implementation has a solid foundation with good design patterns, but has several critical issues that should be addressed before merging to production.
detail mapping #240 2026-02-19 We've implemented pattern 7 as requested, which involves adding a syncstatus column to track the synchronization state of audit records with Elasticsearch. This approach provides reliability by tracking the sync...
Implementation Progress Report #240 2026-02-19 The audit trail infrastructure has been partially implemented with core versioning, database models, and API integration. The foundation is solid but needs external storage integration and UI exposure.
Service Manager — Audit Logs UI & API #240 2026-02-16 This plan covers AUDIT logs only. The system/logs route will expose the auditdocuments table to service managers.
Factor-resolution expression indexes #2404 2026-08-26 The submodule listing's per-row factor lookup scanned the det's whole factor set for every page row (2.08 ms/row over 1219 candidates). Ships 8 partial expression indexes on factors — one per distinct handler kind_field — measured 10× per evaluation, 664 kB total. The 14 s incident itself was stale statistics + cold cache, not the subquery; the structural page-first fix is a follow-up.
Page-first pagination for the submodule listing #2404 2026-08-27 get_submodule_data joined per-row factor resolution and the module's whole emission aggregate before applying LIMIT, so a 20-row page paid a module-sized cost (measured 18,620 of 20,912 buffers on discarded rows). When sort/filter read nothing outside data_entries, the repo now resolves the page's ids first and restricts both expensive branches to them, making cost proportional to the page instead of the module. Falls back to the original shape whenever eligibility isn't provable, so results cannot diverge -- only cost can.
243-backoffice-data-management-factors #243 2026-03-11 (see plan content)
Source Tracking Implementation Summary #243 2026-03-24 Both columns are indexed for query performance.
CSV Data Entry Type Resolution Refactoring #243 2026-03-18 Refactored CSV data ingestion to use module-specific category columns instead of the generic dataentrytype column. This makes the CSV ingestion more explicit and type-safe per module.
Implementation Complete: Delete-Before-Insert Pattern for Factors #243 2026-03-25 Successfully implemented the delete-before-insert pattern for CSV factor uploads to ensure idempotency and prevent duplicate factors.
Plan: Job History Panel + Per-Row Sync Progress #243 2026-03-09 4. Add clearJobHistory() action: jobHistory.value = [].
IT focus graph wrong % calculation (#2441) #2441 2026-08-28 The Results IT-focus waffle claimed ~90% IT share for a unit whose real share was ~9%: the persisted stat divided validated IT kg by the validated totals of only the four IT source modules instead of all modules. Renamed the stat to percentage_of_validated_total and made its denominator the validated total across all buckets. Also fixed the IT-focus total big number rounding (1.2 t rendered as "1" while its parts showed 0.8 + 0.4) by formatting it with up to one decimal.
Backoffice CSV download served the wrong file #2442 2026-09-11 Root cause: archive paths keyed by a recyclable job_id collided across DB resets on shared envs; the file endpoint also cached indefinitely. Fixed both, plus made the collision structurally impossible going forward.
Home chart module icons greyed by access, not status or data #2443 2026-08-28 The Home chart's icon axis greyed modules out based on validation status and data presence (a validated module with no stats bucket rendered greyed as 'not validated', and any module stayed greyed until it had data). Reworked the rule: visible ⇔ activated in the backoffice, greyed ⇔ the user lacks view/edit access; module status and stats presence have no influence. Frontend-only.
2445 — Auto-named plan create loses the unique-index race #2445 2026-08-28 Plan names are display metadata, not identity — unique index dropped, allocator replaced by a random 3-char suffix; heavy-cascade splitting inventoried and routed to
2449 — Plan cascades vs the < 1 s request budget #2449 2026-08-28 Measured stage traces show the planner cascades are fast today — monitoring hygiene and a cheap fan-out trim ship now; the purge job is specified but deferred behind explicit triggers
Unit-specific CSV/API uploads are user-owned — inline edit and delete restored #2453 2026-08-28 #951 bucketed every ingested row into the locked IMPORTED branch. The real line is per-year vs unit-specific: a CSV/API upload into one's OWN module carries carbon_report_module_id and is the operator's own data, so it gets the same edit/delete rights as manual entry. Adds CSV_MODULE_UNIT_SPECIFIC (2) and API_MODULE_UNIT_SPECIFIC (4) to the user branch, backend and frontend.
Tableau empty-transform: prove coherence before treating as routine #2457 2026-08-28 _finalize_empty_transform decided routine (delete-and-replace) vs. genuine error (raise, delete nothing) purely from the names of fired drop reasons, never the values behind them. A casing or date-format drift upstream could tag every row under an 'expected' reason and silently wipe a year's data. Adds a per-provider coherence check that must also hold before trusting the routine path.
Filters #246 2026-03-09 (see plan content)
Optimization #246 2026-02-26 Now I have a much better understanding of the existing seed data infrastructure! Looking at the existing seed files in backend/app/seed/, I can see there's already a well-structured approach for generating test data...
Server-side FTE decimals and activation lock (#2464) #2464 2026-09-24 Two frontend-only guardrails (#2318 FTE one-decimal cap, #2146 activation lock on an open year) had no backend twin, so a direct PATCH, an HR import or a stale tab bypassed them. Both now enforced server-side, with regression tests.
CHANGELOG.md duplication on every dev -> stage merge (#2466) #2466 2026-08-28 changelog.yml (dev -> stage) regenerates CHANGELOG.md via conventional-changelog, which always diffs from the latest git tag. Tags are only cut later, at the stage -> main promotion (release-please.yml), so every dev -> stage run before that tag exists re-diffed from the same stale tag and prepended a fresh overlapping block on top of whatever the previous untagged run already wrote (recurrence of #2352). Fixed by trimming CHANGELOG.md back to the last actually-tagged entry before each regenerate, instead of tagging earlier or dropping the dev -> stage step.
2483 — Explorer/Calculator get-or-create races 500 #2483 2026-08-28 Guard explorer/calculator get-or-create with a SAVEPOINT so the race loser returns the winner instead of a 500
2487 — PUT singletons for explore/calculator provisioning #2487 2026-08-28 Replace the explorer's GET-404-POST provisioning pair with one idempotent PUT (a new ExploreProvisioningWorkflow, ADR-014's first workflow); delete the lazy Calculator-project get-or-create from CarbonReportService.create/bulk_upsert and provision it explicitly in unit_sync
2497 — chunk-load-reload e2e: extra error toast alongside the reload prompt #2497 2026-08-28 chunk-load-reload e2e fails because the unmocked session check's real 404 toasts alongside the reload dialog — mock the session and wait for app readiness before firing the synthetic rejection
2498 — research-facilities taxonomy-unavailable dropdown never opens (e2e) #2498 2026-08-28 taxonomy 404 left the research-facilities select silently empty (aria-expanded but zero options, no message) — surface it as a visible load error in the shared kind/subkind composable, and rewrite the stale e2e test that pinned pre-#2391 fallback behaviour
2501 — building change on a room row must yield an incomplete row, not a 422 #2501 2026-08-28 changing building_name on a buildings room row 422s — the kind-change clear nulls room_type and the
Results Page — Single Summary Endpoint #252 2026-02-20 Currently the page hardcodes most values ("37'250", "8.2", "-11.3%"). The existing GET /unit/{unitid}/{year}/totals endpoint is incomplete — it only sums equipment and doesn't return FTE or per-module breakdown. We...
Carbon Report Module — Stats Column Population #252 2026-03-03 The carbonreportmodules table has a stats JSON column that was never populated. Every time the frontend needs emission breakdowns or scope totals for a module, it triggers live aggregation queries across...
Chart Data Endpoints — Implementation Plan #252 2026-02-18 The results page charts (ModuleCarbonFootprintChart and CarbonFootPrintPerPersonChart) previously used hardcoded mock data. A backend endpoint now returns real aggregated emission data so both charts display actual...
Emission Computation — Data Reference #252 2026-03-02 Some formula inputs are not in dataentry.data and must be derived before.
Emission Type Migration — Implementation Plan #252 2026-03-02 We are replacing the old EmissionTypeEnum (int-based, flat) with a new.
Test Plan: 252 Endpoints #252 2026-02-20 Business logic calculations are inline in the getresultssummary and getvalidatedtotals routes. They cannot be unit tested. They must first be extracted into pure functions.
Validated Modules Totals & Results Summary #252 2026-02-20 The existing aggregation endpoints only handle equipment and don't filter by validation status. Headcount has no DataEntryEmission records — FTE is stored in DataEntry.data["fte"] and must be queried separately via...
Streamline number formatting errors #2520 2026-09-11 Equipment usage-hours fields now go through the same dot-decimal format check as every other numeric field, in the form and in the inline table, instead of coercing a comma value to NaN or forwarding it to the backend.
Denormalize the module key onto data_entry_emissions (#2527) #2527 2026-08-30 Denormalize the immutable join keys carbon_report_module_id and data_entry_type_id onto data_entry_emissions, plus a covering index (carbon_report_module_id, data_entry_type_id, data_entry_id) INCLUDE (kg_co2eq, emission_type_id, primary_factor_id, scope), so the kg_co2eq sort aggregate becomes one contiguous index range scan instead of a nested loop probing ~6 scattered pages per entry.
2539 — Role sync deprovisioning story #2539 2026-09-02 Give role sync a real deprovisioning story: two-strikes confirmation before an empty response ever revokes, backoff on repeated skips, a configurable sync TTL, and an admin-only force-revoke endpoint.
Release the DB pool on shutdown, and make real connection counts visible #2566 2026-08-31 Pods exited without closing their connection pool, orphaning ~10-20 Postgres backends per rollout until TCP keepalives expired; dev filled max_connections and 500'd. Dispose the engine in the lifespan, size the pool from measurement, and export the two series that would have shown it coming.
Let the workspace guard handle a refused unit instead of the HTTP layer #2570 2026-08-31 GET /units/{id} answering 200 means the user may read a unit, not enter its workspace. The workspace-home call is the real authorization boundary and was issued without skipErrorCodes, so a 403 there triggered the global toast + hard redirect to /unauthorized before the guard's own soft redirect could run.
Count DB pool timeouts and connection failures instead of inferring them #2572 2026-08-31 Both DB failure modes were alerted on through gauge thresholds standing in for countable events, so a burst that started and recovered inside the alert window never fired. Two OTel counters make increase(...) > 0 possible, and DB_POOL_TIMEOUT drops 30 → 5 so a timeout costs a user 5 seconds instead of 30.
Normalize pre-existing entry data payloads (audited operator pass) #2592 2026-09-08 Plan for #2592: put the join keys stored in data_entries.data in the same format as the shared field types from #2585. Audited on prod and stage: nothing a factor lookup uses differs, so no migration ships; an operator script (dry run, explicit --apply) replaces it. Decided 2026-09-08 in PR #2585.
Planner: mirror PlanPolicy in the delete gate, and stop the permission snapshot going stale #2607 2026-09-01 Closes the two ways the planner delete button could be enabled while the backend answers 403: canDeletePlan now resolves the permission breadth and applies PlanPolicy.can_delete's rule (global, else creator-only) instead of accepting any delete grant, and the workspace-home aggregate carries freshly computed permissions so the SPA stops gating affordances on the login-time snapshot. No schema change, no product-rule change.
2613 — backend-owned labels for enum-key classification fields #2613 2026-09-01 Seed en+fr labels for enum-key classification values, serve every label backend-side, delete the frontend $te/optionLabelKey/optionLabelPrefix fallback machinery.
2623 — Alert on role sync's suspicious-empty guard #2623 2026-09-02 Count role sync's suspicious-empty and provider-unavailable skips as an OTel metric, and alert on it in openshift-app-config — the guard from #2531/#2538 logged at ERROR but nothing was ever notified.
Lighthouse CI: Frontend Route Coverage #264 2026-04-02 Lighthouse CI previously only audited the login page because all.
2651 — Planner/Explorer requested the wrong year for taxonomy #2651 2026-09-03 Planner reports without a reference year no longer resolve their factor year from "the unit's latest Calculator report year" — an ungated tier that could silently land on the current, unpublished year. Removed in favor of the N-1/N-2 fallback
Planner research facilities: validate the planned use inline instead of surfacing the raw backend error #2653 2026-09-11 A planned use above 100 for a %-metered platform reached the backend, which rejected it with a pydantic ValueError the grid displayed verbatim. The Planner grid now validates each row against the same per-unit bounds table the Calculator form and the backend use, shows the standard 'Must be at most 100' message under the field, and never sends the invalid value.
SSE streams pinned one DB connection each; the 'connection lost' toast was the pool timing out #2654 2026-09-03 Both job/pipeline SSE endpoints authenticated through get_current_user, whose get_db session FastAPI releases only after the response is fully sent. For a stream that is its end, so every open stream held a pooled connection in an open transaction until the job finished. On stage the pod cap (2+15) filled, new stream requests timed out after 5 s, and the browser reported 'Connection to sync job lost'. The streams now resolve the user in a short-lived session.
2656 — Explorer is not year-agnostic #2656 2026-09-03 Explorer's Simulator Explore sandbox is no longer keyed by year or refreshed on a 24h TTL. "Start an exploration" (a page mount or refresh alike) always POSTs a brand-new sandbox; the backend deletes the caller's older ones in the background right after. Factor resolution for Explore entries now always uses the latest started year (N-1, fallback N-2, else a loud error) instead of the sandbox's own year. Addendum — the resolved factor year is now exposed to the frontend and actually wired into Explorer's/Planner's option lookups (#2631,
Backend suite red: three unrelated causes (#2684) #2684 2026-09-09 The backend suite is red in three unrelated ways: a stale config test left behind by the #1153 revert (red on CI now), backend/.env leaking into every pytest process through an lru_cache ordering bug (11 integration tests ran against real EPFL S3), and a CSV fixture #2253 never committed. Three independent fixes, shipped in that order.
DB queries queue at the DBaaS PgBouncer: query_wait_timeout waves #2689 2026-10-09 Dev lost two hours on 2026-09-08 to psycopg ProtocolViolation query_wait_timeout waves, each 120 s long. The error is PgBouncer's client-wait timeout: DBaaS bounces dev, stage and prod, and the server-side connection count plateaus at ~40 in dev. Our SQLAlchemy pool was never the wall (no QueuePool limit error), it was 17 coroutines each holding a slot while queued at the bouncer. Shipped: get_current_user hands its connection back before any route body runs (the general form of #2654), the DB health poller no longer freezes /ready for two minutes per wave, the orphan-poller log stops spamming, and both pods move to 5+50 in openshift-app-config.
Worker hands running jobs back on SIGTERM; how it scales #2696 2026-09-23 A worker rollout killed the running job at the default 30 s grace and left its row RUNNING under a dead pod until the stale sweep 5 min later. The lifespan now cancels in-flight jobs first; each stops its handler, rolls back and releases its row (NOT_STARTED, unlocked, attempts kept) so the next poller tick re-dispatches it. The worker Deployment gets a 60 s grace period. Replica count per environment stays gated on the DB budget; automatic scaling would key on the NOT_STARTED backlog, not CPU, and is capped by DB slots per worker.
Headcount — student form subtitle shows the raw key (#2699) #2699 2026-09-15 The Headcount student form ("Add student FTE") rendered the raw key headcount-student-form-subtitle: the subtitle text was removed from the translations on 2026-07-07 but the submodule config still had hasFormSubtitle: true. The flag is now off; the form has no subtitle.
Issue 2702 — References upload through the shared import dialog; upload cards accept a dropped CSV #2702 2026-09-09 The references card in the backoffice configuration page was a 480-line fork of the shared UploadCard with a hidden file input that uploaded on selection, bypassing the import dialog's overwrite warning and the year-sync guard. It is now a thin wrapper routed through the same dialog as Factors and Data. Every upload card is also a drop zone: dropping a CSV uploads it straight away through the dialog's upload path, with no modal. Enter after picking a file in the dialog triggers Save.
2706 — One path for a module's stats #2706 2026-09-11 Module stats had two implementations: a JSON persisted by the pipeline and a live aggregate recomputed on every module-detail GET. The GET now reads the persisted column; the per-viewer planner-snapshot filter and the headcount FTE chart maps are persisted alongside so nothing is re-aggregated at read time.
Building-room reference validation (#2716) #2716 2026-09-11 Close every silent-zero path in the building-room reference pipeline: the reference pack is validated as a whole at import (empty cells, duplicate rooms, buildings without factors), uploads naming a room with no surface are rejected, room names match with spaces ignored and persist in the reference spelling (#2268), a building with no factor row fails at compute instead of producing no rows, and the unvalidated seed_building_rooms importer is removed.
Planner equipment: global percentage in the Detailed per Year sections #2749 2026-09-18 The equipment 'Planning mode' toggle (manual entry per line vs one global percentage of the reference year, #1981) existed only in the Project Grant section. It is now offered in every Detailed per Year section too; the backend route behind it no longer rejects non-grant plan reports. Budgets stay grant-only. Follow-up: global mode at 0% is now the prefill default, so equipment no longer copies every reference line.
Explorer — travel not added in the table (#2752) #2752 2026-09-14 In an Explore sandbox a standard user's new flight/train trip was created but never shown: the professional-travel own-rows read filter (traveler id == caller's institutional id) hid it, because the roster-less Explorer stamps the "other traveler" sentinel. Explore sandboxes are private per user already, so the own-rows filter is now skipped for them.
recompute-stats never reached Simulator Plan reports (#2775) #2775 2026-09-14 Scope the recompute-stats and stale-stats queries on the effective factor year (COALESCE(reference_year, year)) instead of carbon_reports.year, so Simulator Plan reports fold into their baseline year's aggregation scope instead of raising a scope for their planning year — which has no factors, is never selectable in the operator UI, and left every plan's stats permanently un-recomputed. The aggregation handler's module slice is widened for the admin trigger only, so the recalc-chained path cannot bump validated plans to IN_PROGRESS.
2818 — Layout breaks on smaller screens #2818 2026-09-16 Module page and header fit laptop-sized viewports instead of clipping at 1320px.
Asset caching: hashed immutable, unhashed revalidated, public/ assets through Vite #2835 2026-09-17 Travel map broke because maplibre's ES-module worker was served as application/octet-stream and, under 1.4.13, cached as immutable in that broken state. Fix in three layers: register mjs/csv MIME types and bundle the worker with Vite (#2837, #2838); nginx gives immutable only to Vite-hashed names and revalidates everything else; templates, fonts, favicons, logo and login image move from public/ to src/assets so a replaced file gets a new URL. Also the cache side of #2812 (stale CSV template).
Downloads go through fetch + Blob so they are observable and fail loudly #2840 2026-09-17 Six places built a hidden navigation by hand: the request never showed in DevTools and a 401/404 died in the browser's download shelf without reaching the app or Sentry. One downloadBlob helper plus a fetchFile API module replace them; the ?d=true Safari workaround goes because object URLs honor a.download everywhere. Follow-up from #2835.
Undecryptable stored file answers an explicit 500 #2842 2026-09-17 GET /api/v1/files/{path} logged 'Error retrieving file: ' with nothing after the colon and returned a blank 500 when the object could not be decrypted: cryptography's InvalidToken has an empty str(), the handler logged only str(e), and the exception never reached the span. The handler now catches InvalidToken by name, logs with traceback and path, records the exception on the span, and answers a detail that names the cause. Cleanup of the objects themselves is an ops task.
2845 — Taxonomy cache broadcast blocked by NetworkPolicy #2845 2026-09-17 The chart's default-deny NetworkPolicy also dropped worker→backend and backend→backend traffic, so the taxonomy-cache broadcast (#2258 follow-up) never reached a pod on dev/stage. One extra policy admits pod-to-pod ingress to the backend port inside the release.
Lock the recalc step only #2847 2026-09-17 Unit-specific uploads of the same module ran head to tail because the ingest took the per-module advisory lock before inserting. The ingest is append-only, so it now takes the shared factor gate only; the chained recalc keeps the module lock and is the one step that serialises. To keep that safe, the scoped recalc dedup stops collapsing onto a RUNNING sibling.
Workers view lists job-running pods only #2853 2026-09-17 The workers view listed API pods because every pod heartbeats and the row carried no role. A runs_jobs flag on the heartbeat row, set from the pod's dispatch switches, lets GET /v1/sync/workers list job runners only.
Job queue-wait and duration histograms #2854 2026-09-18 Job queue wait and duration were columns nobody queried; two OTel histograms per job_type put p95s in Grafana so pod count and MAX_CONCURRENT_JOBS get set from a week of real use instead of a hand-run SQL.
Consolidate Alembic migrations before go-live #2904 2026-09-22 Collapse the 20 Alembic migrations into one before go-live; dev, stage and prod are dropped and rebuilt from it. Label seeds stay as reference data in the new migration.
2909 — Deploy from gitlab.epfl.ch #2909 2026-11-18 dev, stage and release tags deploy from gitlab.epfl.ch through the EPFL-ENAC/build-push-deploy CI/CD component. dev moved first, when GitHub Actions was billing-locked. stage and tags followed on 2026-11-18, when quay-its.epfl.ch closed to traffic from outside EPFL, and deploy.yml was deleted.
Clean prod logs: probes, collector conflict, policy chatter #2934 2026-09-24 A 24 h prod Loki dump was 97% probe access lines and collector self-warnings; drop the probe lines at the source, delete per-evaluation policy chatter, route every 403 gate through one logged helper, and ship the dump script; the collector fix is config in openshift-app-config.
Single sliding session cookie #2943 2026-09-25 One sliding httponly session cookie renewed inside the auth dependency and audited on renewal replaces the access + refresh pair and the client-driven POST /session; GET /session answers 200 with user null when anonymous. Shipped in one step, no compatibility window, after the maintainer chose to test the cut-over on dev and stage.
2992 — provider change on a cloud row must yield an incomplete row, not a 422 #2992 2026-10-02 an inline provider change on an External Clouds row 422s — the kind-change clear nulls service_type and resolve_clouds fails hard on the missing value, rolling back the edit; fix mirrors
2997 — stale permission patches and a leaked shutdown flag #2997 2026-10-02 24 backend tests failed on dev — 23 integration tests still monkeypatched is_permitted on route modules that
3000 — full apk upgrade in the runtime images, urllib3 2.8.0 #3000 2026-10-02 Trivy blocked the dev deploy on HIGH CVEs — pcre2 10.48 in the nginx-alpine docs/frontend images, urllib3 2.7.0 in the backend venv; the three runtime stages now run a plain apk upgrade --no-cache instead of naming one package per CVE, and uv.lock bumps urllib3 to 2.8.0
3099 — type-check and frontend tests after the 2026-10-19 deps update #3099 2026-10-19 After the 2026-10-19 dependency consolidation, make type-check failed: ty 0.0.89 flags narrowed overrides, and vue-tsc 3.3.13 has a template codegen bug. The frontend suites also found no tests, because @playwright/test 1.64 sat next to ct-vue 1.62.1. Fixed in code without pins: backend overrides keep the base type, template closures go, and @playwright/test is dropped so ct-vue is the single Playwright version.
310-a — Pod Safety + Atomic Claim #310-a 2026-05-06 Pod-safe atomic claim mechanics for FastAPI background tasks under multi-pod deployment.
310-b — Factor Pipeline + Unit Sync Tracking #310-b 2026-05-06 Bulk-path fixes: factor ingest semantics, recalc enqueue, and unit-sync tracking on Path 2.
310 — Backoffice recalculation upon factor updates #310 2026-05-06 Recalculate downstream emission data when operators update factors via the backoffice.
310-c — DAG + Handler Registry + Observability #310-c 2026-05-06 DAG-driven handler registry with observability columns replacing ad-hoc safety and auto-trigger code.
310-d Architecture follow-ups (#1062 / #1063 / #1064) #310-d 2026-05-29 Three architectural follow-ups staged from the 310-D landing — passive-monitoring backstop, unified pipeline state store, and DedupConfig generalisation. Each is tracked on a dedicated GitHub issue and lands in the post-merge fix batch.
310-d Frontend — Stale-Stats UX + Pipeline SSE #310-d 2026-05-07 Render the in-flight bulk-pipeline state on carbon-report module cards, subscribing to GET /sync/pipelines/{id}/stream until the chain finishes.
310-d — Bulk Path Pure Async (Path 2 only) #310-d 2026-05-07 Make the bulk pipeline (Path 2) pure async to match its latency profile; Path 1 untouched.
Plan 310-D Follow-up — Strategy B Rematch + Per-Module ITs #310 2026-05-06 Rematch the FK-link modules (travel, headcount, building embodied energy) plus regression coverage for JSON-link modules including the 1:N building-rooms fan-out.
310-e Async-loop hygiene & pipeline worker hardening #310-e 2026-06-15 Improvement backlog from the 2026-06-15 dev incident: a CSV ingestion job ran in-process on the web event loop, starved /healthz past the 2s liveness timeout, and the pod was restarted mid-job. Captures the shipped mitigation plus the larger refactors it points to — a dedicated pipeline worker, event-loop offloading, and poller/durability sharpening.
310-f Ingestion per-row efficiency #310-f 2026-06-16 The CSV ingest parse/validate loop dominated wall-clock (~34s for 9555 purchase rows on dev). The first cut guessed line-level suspects and they were wrong. A row-loop profiler was added, and the trace pinned 33.4s of 34.5s on resolve — a per-row linear scan of ~23k factor keys in lookup_data_entry_type_by_kind (O(rows × factors)). Fixed by memoising the kind→type inference per (kind, subkind); re-measured 34.5s → 1.7s (20×). A one-time index is a documented follow-up, not needed at current data shape.
310 — Background Pipeline Architecture: Overview #310 2026-05-06 Umbrella plan: redesign FastAPI background workflows for multi-pod safety, retry, and observability.
310 — Post-merge fix batch (review findings + #1062/#1063/#1064) #310 2026-05-07 Parallel worker batch landing the 8 High/Critical findings from docs/code-review/310-overall-review.md plus completing the three architecture follow-ups, all targeting feat/310/dev integration branch.
Plan: Manual emission recalculation endpoint with "needs recalculation" status #310 2026-04-10 no new DB table — derived from existing DataIngestionJob rows.
310 — Plan Review #310 2026-05-06 Code-grounded verification of the 310-* plan series against the actual repository state.
3140 — Make an Accred outage visible to users and ops #3140 2026-10-27 Accred 5xx, 401/403/429 and timeouts become RoleProviderNetworkError: the login callback redirects to a login page that says Accred is down, a new accred.request.failures counter feeds a critical AccredUnavailable alert, and role sync records the outage instead of swallowing it.
3142 — Move train trips off removed bus stops #3142 2026-10-29 Option 1: before the next train recalc, train trips at the 3 bus stops the rebuilt reference drops move to the nearest railway station. Hand-run SQL right after the reference upload, then a train recalc per affected year. No code change.
Recalculate emissions card on the pipeline console (#3153) #3153 2026-10-29 Add a Recalculate emissions card to the pipeline operations console (module, optional data entry type, year, confirm), open both POST /sync/recalculate-emissions endpoints to backoffice.pipeline_operations edit as well as backoffice.configuration edit, and delete the dead per-type recalculate emit chain left behind when #780 removed its button.
3185 — Code Auditor follow-up: drop passlib, refresh certifi #3185 2026-11-05 Code Auditor follow-up. Drop passlib and other forgotten dependencies, move the FastAPI OTel instrumentation to main deps, move the dev-only Faker seeders out of app/, and bump certifi in the docs lock.
Code-scanning sweep (#3185) #3185 2026-11-05 Cleared the open CodeQL and Code Quality findings on dev. Two security alerts, dead code and test hygiene fixed in code; the remaining false positives dismissed.
Security audit job-d0adab0d: remediation #3185 2026-11-10 Remediate the security audit job-d0adab0d: object-level authorization and dependencies first, then boot, session and policy hardening, CSV neutralization, CI hygiene and temporary-upload scoping.
3271: registry scan to code scanning #3271 2026-11-16 Registry scan findings become code scanning alerts (build-push-deploy-action v4.0.0), the docs image joins the scan, and dnspython moves to 2.9.2.
Zensical, security commitment and Pages from the release image #3333 2026-11-20 SECURITY.md states the CodeAuditor commitment. The docs move to Zensical. GitHub Pages deploys from the release's docs image.
/me Performance Optimization: Decoupling Role Sync and Frontend Refresh Strategy #334 2026-05-06 Drop /me latency from ~1s to ~8ms by moving role sync trigger from /me to /refresh.
Frontend Error Monitoring via Self-Hosted GlitchTip #344 2026-05-19 Wire @sentry/vue to EPFL's self-hosted GlitchTip with runtime DSN injection under a read-only root filesystem.
Implementation Plan: Unify CSV Upload Ingestion #370 2026-02-04 Unify CSV upload + API sync under one job-creation pipeline that stores file paths in job meta, dispatches providers via the factory, and exposes job status via SSE.
Plan: Travel — Link Traveler Selection to Headcount (Issue #372) #372 2026-03-10 Travel entries currently store travelername as a free-text string and travelerid.
Feature Overview (Why?) #384 2026-04-02 In the calculator a carbon footprint appears in red if crossing a certain threshold, as defined in Data mgt.
Simulator Module #404 2026-07-16 Standalone module letting any authenticated user project a research project's (or single category's) carbon footprint by reusing reference-year annual unit data, without writing back to the data-entry modules.
Security: authentication & integration hardening #458 2026-05-27 Pin the trust boundaries of the OAuth/JWT auth path with regression tests, fix eleven of twelve findings (F6 logout-denylist deferred to a follow-up — needs server-side store), centralize the JWT→user resolution path, and document the contracts.
Backoffice ACCRED affiliation scoping #459 2026-05-26 Scope backoffice.* permissions by ACCRED-provided affiliation so each backoffice manager sees only their sub-perimeter.
PDF Export for Backoffice Reporting — Implementation Plan #462 2026-08-28
PRD 2 — Results CSV (results.csv) #504 2026-03-31 👉 If byemissiontype is missing → STOP (data issue).
Plan: Headcount user_institutional_id — Uniqueness Validation + Institution Label #518 2026-03-10 The headcount module stores a userinstitutionalid (SCIPER at EPFL) per member entry.
Implementation Plan: Headcount Data Validation V2 (Issue #518) #518 2026-03-18 This plan covers five distinct changes to the Headcount module, derived from the updated.
PRD 3 — Detailed CSVs (*_data.csv) #589 2026-03-31 Provide raw, traceable, and schema-specific data for each data entry type.
CSV upload: surface per-row validation errors with readable messages #659 2026-07-09 Turn raw pydantic ValidationError dumps into readable field-level messages, and fix the upload-card UI which currently never renders them.
Homepage tooltip: webapp's own Lighthouse/ecoindex footprint #673 2026-07-07 Surface the nightly Lighthouse/ecoindex audit of the homepage as a small info tooltip in the homepage welcome header.
Refactoring: Data Management Upload Components #780 2026-05-06 Deduplicate ~200 lines of upload card UI between ModuleConfig.vue and SubmoduleConfig.vue.
Compare Years: aggregate across combined units + reword target box #834 2026-08-21 Replace the single-unit multi-year-report-stats endpoint with GET /merged/multi-year-report-stats (unit_ids query, allow-list auth factored out of the year-scoped helper), wire combined units from ResultsPage through CompareYearsDialog, and reword the target-gap KPI to the approved 'Reduction needed to reach {year} target' phrasing with a template-side minus sign.
Remove emit_per_factor — always emit one emission per factor #840 2026-05-06 Eliminate the emit_per_factor branch in prepare_create(); always emit one DataEntryEmission per factor.
Root-Level Rollup DataEntryEmission Rows #840 2026-05-06 Add root-level rollup rows on DataEntryEmission to support aggregate views.
Replace distance_km / weight_kg with polymorphic additional_value on data_entry_emission #840 2026-05-06 Replace the two typed columns with a polymorphic additional_value field on data_entry_emission.
Issue 856 — Simulation Explore Flow #856 2026-05-06
Issue 857 — Back-office: Open year for users (functional, end-to-end) #857 2026-05-12 Consolidate the back-office data-management page year-lifecycle: single-click create+sync, real pipeline observability, hide unstarted years from end-users, and wire the 'Open year for users' button.
Refactor Seed Factors to Use Ingestion Machinery #859 2026-05-06 Replace duplicated factor-seed parsing with a LocalFactorCSVProvider that reuses the ingestion pipeline.
Refactor seed_generic_data_entries to Use Ingestion Machinery #859 2026-05-06 Replace duplicated data-entries seed logic with LocalDataEntryCSVProvider on process_csv_in_batches.
#860 — Dev docs refresh & MkDocs restructure #860 2026-05-06 20-PR refresh of developer documentation across 4 work-streams (A/B/C/F) consolidating, condensing, and structurally reorganizing docs/, with new frontmatter conventions, auto-indexing, ADR backfill, and mermaid-diagrammed architecture.
Backoffice affiliation scope by unit subtree (cf-anchored, intersection clamp) #862 2026-06-03 Resolve backoffice.reporting affiliation scope from a single ACCRED unit cf (any level) into a descendant unit-id set via path_institutional_code, and clamp every backoffice query by intersecting the scope set with the endpoint filter set. Replaces path_name substring matching and list-narrowing.
Update backoffice permissions (page-driven model) #862 2026-06-02 Replace the system.users super-admin gate with one permission per backoffice page (configuration, pipeline_operations, logs, ui_texts), scope only backoffice.reporting by affiliation, and move the #459 affiliation anchor to backoffice.reporting.
CSRF: Origin/Sec-Fetch-Site enforcement — implementation plan #89 2026-08-26 Fail-closed request-origin middleware to close the *.epfl.ch sibling-subdomain CSRF gap. Compact version; full rationale and endpoint audit in the long-form plan.
Fix ModuleUnitSpecificCSVProvider KeyError + Add CSV Fixture Tests #898 2026-05-06 Resolve KeyError on factor_id_to_factor in ModuleUnitSpecificCSVProvider; add fixture-based tests.
Workspace shows only level-4 units #930 2026-05-22 Backend /users/units endpoint hardcodes level == 4 so workspace only ever sees actual lab-level units, not EPFL or affiliation-level ancestors.
Tighten unit-scoped permission gates #977 2026-05-22 Close three pre-existing permission soft-spots: 2 carbon_report endpoints + 3 unit_results endpoints get unit gating; remove the coarse modules.* fallback in files.py and data_sync.py.
996 — _process_row factor counter always reports rows_without_factors #996 2026-09-02 BaseCSVProvider._process_row hardcodes the factor tuple slot to None; dropping the unreliable rows_with_factors/rows_without_factors counter (Option C) rather than inventing a signal that doesn't exist yet — and the same dead counter in the Tableau API-ingestion hierarchy, folded into this PR at Guilbert's request.
Docs: single source of truth for auth flow #1318 2026-06-29 Strip duplicated and stale auth narratives from backend and frontend overview docs. The architecture/04-auth-flow.md doc is the canonical source; everything else cross-links.
Bootstrap year configurations from INPUT_DATA #1879 2026-07-23 make bootstrap-years replays the whole post-db-drop backoffice click-path — year configuration, unit sync, every factor CSV, the three reference CSVs, and the reduction-objective files + goals — from backend/INPUT_DATA, for 2025 and 2026. No data entries.
Code-scanning cleanup none (GitHub code-scanning sweep) 2026-11-05 Resolve all 84 open CodeQL code-scanning alerts — 2 security fixes with regression tests, ~30 mechanical cleanups, and triage dismissals for false positives / deliberate patterns.
Emission Pipeline Flow 2026-05-06 Reference for the emission compute pipeline AND for how factor changes propagate to existing emissions (Plan 310-D rematch).
Stop persisting computed fields to DataEntry.data 2026-05-04 Stop persisting computed fields back into DataEntry.data; recompute on read instead. Single-PR refactor delivered during sprint 10; no tracking issue.
Authentication Hardening Implementation Summary 2026-03-20 This document summarizes the implementation of the "Harden Authentication Identity Resolution" PRD to fix a security incident where test users could collide with real user IDs.
Fix: sub_class field disabled state in ModuleForm 2026-03-12 When an equipmentclass (e.g. "3D printer") has an empty subclass array from /api/v1/factors/{id}/class-subclass-map, the subclass field should be hidden/disabled. This worked in ModuleTable (via...
Load the brand font from enac-it-fonts.epfl.ch 2026-10-28 Load Suisse Int'l from the ENAC font server shared by ENAC apps instead of from the app's own build output.
Frontend error reporting — minimal GlitchTip client (drop @sentry/vue) #1569 2026-06-18 Replaces the @sentry/vue SDK with a ~2 KB dependency-free GlitchTip-compatible reporter (src/utils/glitchtip.ts). Captures Vue component, router, global, DOM-event, unhandled-rejection and ky 5xx errors. Fixes the two protocol bugs (missing envelope item header, missing DSN public key) that stopped events ingesting, and documents that the DSN must live in .env.local.
Drop Material Icons webfonts in favour of tree-shaken SVG icons 2026-09-10 Replace the two Quasar Material Icons webfonts (~278 KiB of woff2, render-/LCP-blocking, font-display:block FOIT) with in-component SVG imports from @quasar/extras and the svg-material-icons icon set, so only the icons actually used are bundled and no icon webfont is shipped.
Homepage CO₂ first-load badge #673 2026-09-02 Build-time CO₂ first-load badge in the homepage calculator header
Ops runbook and tools documentation none 2026-09-07 Day-to-day operations documentation — an incident landing page, a tools-and-access page, and a rewritten infra overview with every stale ops claim removed.
Explicit RoleScope: own / unit / global #862 2026-06-02 Replace the implicit role→scope coupling with a discriminated-union Scope (global/unit/own/affiliation) so the own↔unit boundary is explicit in both the schema and the permission key, fixing unit-level capability leaking to standard users.
Stats buckets: replace emission-breakdown with persisted report stats #841 2026-07-14

In progress (30)

Title Issue Last updated Summary
Data-validation audit and hardening — backend as single source of truth #1489 2026-09-18 Slice 1 of #1489: a systematic audit of every documented data rule (back-office doc site, data-description pages) against actual backend enforcement, frontend enforcement, and test coverage, across the four ingestion paths (back-office reference/factor CSVs, entry CSVs, direct form input, reduction-objective CSVs). Every claim was re-verified by probing the DTOs and replaying fixtures. Mismatches are classified code-gap or doc-stale (the latter handed to the data manager, not fixed in code), and the audit ends in an ordered set of small follow-up PRs, starting with the #1545 reference-CSV silent-wipe hole. The cross-repo doc-change automation bonus is parked.
ModuleTable.vue decomposition #1557 2026-08-28 Break the ~2017-line shared Calculator/planner data-entry table into composables + sub-components, each under the repo's 500-line limit, with no behavior change.
Planner Purchases: multi-currency input with automated EUR conversion #1918 2026-08-19 The planner Purchases section (per-category totals and global budget) gets a section-level currency selector mirroring the one already visible in the Calculator purchases module. Stored amounts remain EUR-only (amount_eur, no migration): the backend converts a non-EUR input to EUR on write via the existing ECB ExchangeRatesService, and the frontend converts the stored EUR value back into the selected currency for display through a new read-only exchange-rates endpoint. Planner purchase module only; the latent Calculator factor-currency bugs are documented as follow-ups.
CSV templates: inline instructions as # comment lines #2026 2026-09-01 The data manager delivered a new template pack (SharePoint _DATA_INPUT/templates, 2026-08-24) that carries filling instructions as ordinary CSV rows, plus three defects that fail on upload: two files in latin-1, US-format example dates, and four filenames the app does not serve. Teach the shared reader to skip #-prefixed lines, then ship the pack with the instructions rewritten as comments and the defects fixed.
Optimize Pipeline/Report Performance — what remains after the v3 investigation #2049 2026-08-26 Updated 2026-08-26. The v3 trace investigation is finished and its alerting half shipped as #1402. The caching, batching, upload-auth and http.route work has since merged (#2280, #2266, #2265), and #2220 is closed. What is left: confirm the win with one measurement, the OTel instrumentation tax, the remaining ops/alerting items, and three gated changes. Ten claims are corrected here - read those first; two of them retire proposals this plan itself made.
Ceiling-scale performance fixtures (#2161) #2161 2026-08-28 Turn #2161's real per-data_entry_type ceiling estimates into a reusable Postgres fixture (one unit seeded at worst-case scale via the production recalc path, so factor resolution is real) and a systematically -discovered suite that hits every GET endpoint under app/api/v1 against it, asserting a 200ms ceiling-scale budget.
Table pagination + sort matrix (#2295 follow-up) #2295 2026-08-29 Exhaustive latency matrix for the table endpoint GET /v1/carbon-reports/{id}/modules/{module}/{sub}: every module × submodule × limit {20,100,500,1000} × every sortable column × asc/desc, plus filter search, deep pagination, item GETs and the module chart companions, against a full-ceiling unit; sorted big pages folded into the concurrent ModuleReadUser ladder.
Classification label i18n: translation table + generic CSV convention, equipment + purchase #2401 2026-09-01 Adds a (field_name, value, lang) translation table (proposition 2) fed by a generic _ CSV column convention, wires the taxonomy tree builder AND the submodule search filter (#2516) to serve/match localized labels per request. Ships against equipment and purchase; other modules (buildings, headcount, research_facilities, external_cloud_and_ai, process_emissions — per charlottegiseleweil's field list on the issue) pick it up by adding the suffixed CSV column, no further backend code change for the self-labeling shape.
2527 — read-path query counts (items 4, 5, 6, 8, 10) #2527 2026-09-28 Read-path items 4, 5, 6, 8 and 10 of #2527: collapse the per-report loops in the merged modules-stats trio and the workspace-home bundle into grouped queries, cache the year-configuration, and make the SQL-statement budget a per-PR CI gate so the next N+1 fails at review time. Tracing now keeps every request span and SQL spans for 10 % of traces: 36 → 21 ms of CPU per request on dev.
2527 items 1-3 — the three write paths over the 10 s alert bucket #2527 2026-09-10 JobLatencySLOBreach fires on a 10 s bucket; uploads, plan prefill and plan mutations sit above it under load. Per-path verdict from the code: uploads and prefill are queue-plus-CPU, not per-row SQL; plan PATCH/DELETE round trips were already trimmed by #2449 Track B, so its residual is cascade data volume — which fires #2449's own deferred trigger for Track A. One correctness fix (module-scoped emission_recalc children are silently dropped by EMISSION_RECALC_DEDUP) must land before any baseline, because it flatters today's upload numbers.
Concurrent CSV upload move failures (#2528) #2528 2026-09-02 The reported 56%-at-20-parallel rate was void — an artifact of a split file store in the reporter's own test topology, not S3 concurrency (see Correction). A clean re-run at 5 and 20 concurrent uploaders against real S3 (2026-09-02) then measured zero move failures, closing the open question. What survives: the move is unretried, and enacit4r-files 1.4.0 already aligned the two backends on raising — the app's dead falsy-return branch just hasn't been removed yet. Plan: catch the shared exception and retry at the single choke point, shipped now on its own merits; client-pool sizing and retry-mode tuning stay deferred — no measured problem for them to fix. Also surfaced, out of scope: dev-DB/PgBouncer connection contention at 20 concurrent users.
2530 — Rate limiting: router annotations + per-user in-app limiter #2530 2026-08-30 Two-layer rate limiting — OpenShift router annotations for the hard cap, an in-app per-user limiter on expensive endpoints as a per-pod safety valve. Part 1 (unforgeable audit IP) shipped; this covers Part 2.
2531 — Role sync must not read "I got nothing" as "there is nothing" #2531 2026-09-02 Stop background role sync from erasing a user's roles when the role provider answers with an empty list, and make every way of getting that empty answer loud instead of silent.
Equipment: carry forward prior-year usage with factor/default fallback #259 2026-08-18 At bulk ingest, usage hours set in the unit's most recent prior year are carried onto the matching equipment_id (per field, one lookup query per unit/year); a field the prior year left unset keeps tracking the factor suggestion, and when the factor has none either the emission formula falls back to the spec defaults (active 12h, standby 156h per week). No schema or frontend change.
Frontend usage analytics — self-hosted Matomo (ENAC web analytics) #2649 2026-09-03 Adds cookieless Matomo page-view tracking to the Quasar SPA, gated on a per-instance site id. Config follows the existing runtime-injection chain (.env.example → quasar.config.js → runtime.ts → entrypoint.sh → Helm), with APP_MATOMO_URL defaulting to https://enac-webanalytics.epfl.ch/piwik/ and APP_MATOMO_SITE_ID unset by default so dev/CI/unconfigured deploys stay silent. Tracked URLs are normalized to route patterns so unit acronyms never leave the app.
403 on /sync/jobs/{id}/stream for principals on their own MODULE_UNIT_SPECIFIC uploads #2654 2026-09-17 DataIngestionJob.entity_id — the only column the per-job scope check reads — has never been written by any job-creation path since the check shipped (206d9f487, 2026-05-07). Every unit-scoped job resolves to 'no unit', so job_stream_by_id's #1764 fallback 403s every principal on their own upload while the shared per-job gate silently no-ops for everyone. The test matrix missed it because every test mocks the resolver or pre-sets entity_id by hand; nothing exercises create_job. Includes the authz path inventory (6 mechanisms, 9 gate shapes) and 5 consolidation options.
runner.py holds job_session/data_session for the whole job #2700 2026-09-09 Follow-up to #2689: that issue fixed connection-pinning on the auth path and the SSE streams; run_job itself still opens job_session and data_session once and holds both for the job's full lifetime. Part 1 (this plan): job_session's actual gaps are narrower than they looked — base_provider.py already commits it on every progress update — so this closes claim-to-first-commit and last-commit-to-finish_job, no retry-semantics risk. Part 2 (data_session, one commit for the whole batch — job 59 in #2689 was 144k rows over ~70 min) is deferred: it needs an idempotent-resume design and its own reviewed plan, per guardrails on pipeline internals.
MODULE_UNIT_SPECIFIC jobs must not auto-retry after a crash #2700 2026-09-09 Part 2 of #2700, corrected: data_session already commits per INGEST_COPY_BATCH_SIZE batch (since 2026-06-10) — there is no held-open-for-70-minutes transaction to fix. What batching exposed instead: MODULE_UNIT_SPECIFIC uploads are append-only with no delete-before-insert and almost no DB uniqueness constraints, so the stale-job sweep resetting a crashed job to NOT_STARTED for reclaim can re-run the row loop over rows a dead attempt already committed, duplicating them. Fix: MODULE_UNIT_SPECIFIC jobs never auto-retry — a stale one goes straight to FINISHED+ERROR with a status message telling the operator to verify and clean up before re-uploading. Manual recovery (POST /jobs/{id}/recover) is blocked the same way. Accepted by the maintainer: manual row cleanup is the correct resolution, not automated dedup.
2771 — CO₂ formula reference tests #2771 2026-10-29 Reference tests for the kg-CO2-eq formulas: every worked example in the back-office data-description doc runs through the real handler computations and DataEntryEmissionService._apply_formula and must match the domain owner's expected value. Rows where doc and code disagree are pinned as strict xfails and listed for the domain owner; AI and common purchases stay out until the doc and code agree on a formula.
Equipment global percentage: one aggregate line per type instead of rewriting the reference year #2783 2026-09-18 PATCH reference-percentage rewrote every snapshot line one by one (~5,700 queries, 2.3 min for 950 lines). Per the Option B decision in #2749, it now upserts one aggregate line per equipment type, reading the reference total straight from the module's already-computed stats. No schema change, no data backfill.
SSE streams end cleanly on DB errors #2956 2026-09-28 Follow-ups from #2946/#2954. A DB error in an SSE stream's poll loop surfaced as Starlette's 'response already started' RuntimeError, with the real error only in cause. Both SSE streams now log it and end: the job stream with its existing 'Job not found' event shape, the pipeline stream silently so native EventSource retry reconnects.
Institution values, assets and theme from the deployment (#3244) #3244 2026-11-18 Institution name, texts, documentation links, logo, fonts and color theme come from the deployment. The image has neutral default assets, so a deployment gives only its values and a theme.conf. The code has no EPFL default and the texts do not name Accred. A CI check finds hard-coded EPFL, Accred or Entra.
3250: install page and Helm example #3250 2026-11-18 First part of the deployment guide. A correct chart README, example values and Secrets that the lint renders, a generated values reference and an install page.
First screen: record of the explorations (#3298) #3298 2026-11-17 Record of each idea explored for the first screen on 2026-11-17: ten parallel explorations, the logo attempts, the architecture options, and the measurements. It says what was kept, what was dropped, and what to look at again.
Fast first screen: login page and first paint (#3298) #3298 2026-11-17 The login page paints its title from index.html and loads its code in parallel. The first screen needs 3 files in place of 45. Lighthouse mobile goes from 86 to 100. The signed-in pages are Part 2 and not started.
Reuse EmbodiedEnergyWorkflow logic for CSV/bulk building ingestion #878 2026-08-17 CSV/bulk building-room ingestion never derives building_embodied_energy rows (#878) — only the single-item CRUD API path does, via EmbodiedEnergyWorkflow.post_create/update/delete. A prior attempt (PR #2121) fixed the symptom but reimplemented the derivation logic inside BaseCSVProvider instead of reusing the workflow. This plan extracts the shared derivation logic (fixing a latent bug where it silently depends on a frontend-only field), keeps #2121's simpler synchronous/full-replace placement, and reuses it from both the CRUD and CSV paths.
Embodied energy: room_name-keyed derived entries + reconciliation #878 2026-08-18 Refactors building_embodied_energy derived entries to persist only room_name — building_name and surface resolve from BuildingRoom reference data at compute/read time — and rewrites EmbodiedEnergyWorkflow as an idempotent reconcile of the derived-entry multiset against the module's building parents. CSV derivation moves from a job-end read-back to per-batch creation.
939 — Travel from/to search: paging, word matching, trigram indexes #939 2026-10-28 GET /locations/search honours page on a total order and matches word by word, ignoring accents and separators. Each word is served by a trigram GIN index on search_text(column). The travel from/to selects load the next page by infinite scroll.
Edit rights per permissions on each dataset — hardcoded provenance-keyed data-entry permissions #951 2026-08-28 Enforce, per module/submodule, which fields of BackOffice-imported rows vs user-added rows may be edited and whether a row is deletable — keyed on DataEntry.source provenance, matrix hardcoded in code (no DB table, no backoffice UI), enforced in the module mutation workflow and surfaced to the frontend via a per-submodule policy object.
Pipeline-debug — living TODO 2026-05-20

Abandoned (38)

Title Issue Last updated Summary
Provider (tenant) isolation for factors, pipelines, connectors, locations, building_rooms #1832 2026-09-10 Add a provider column (user_provider_enum) to every table that still leaks across tenants, scope all reads/writes like year_configuration does, and backfill by duplicating reference data per provider so TEST starts with a full mirror of ACCRED.
Backoffice: Copy factors from previous year #740 2026-07-13 Add a backoffice action to duplicate a prior year's factor rows into a newly opened year, bulk or per-submodule, instead of forcing a full CSV re-upload.
Professional Travel Module - Implementation Summary #171 2026-01-15 4. Data Entry (Ch 7-8): Form + CO2 calculations.
Plan: Buildings Module (Task 173) #173 2026-02-19 The Buildings module (infrastructure, moduletypeid=3) exists only as a skeleton: enum stubs, a minimal frontend config, and is blocked behind forbiddenModules in the UI. This plan implements it fully per the success...
Permissions-Based Access Control Implementation #221 2026-01-14 Add to backend/app/services/authorizationservice.py (new file).
Cache the per-request user lookup #2278 2026-08-27 Investigation of #2049 item C1. Every authenticated request pays an uncached SELECT users (15.2 ms idle, 71.7 ms burst) whose result carries the roles that permissions are calculated from, so caching it is a permission-scoping change and gated on maintainer review. Recommendation is DEFER: n=2 measurement, #2049 A1 unanswered, and the burst that produces the 71.7 ms figure is already being dismantled by #2275 and B2. If approved anyway, the design is a 30 s process-local TTL on an immutable snapshot, TTL-only with no eviction hooks and no cross-pod broadcast, failing closed on miss.
PRD: Year Configuration & Data Management #244 2026-03-31 Centralize annual administrative settings, emission thresholds, uncertainty levels, and institutional reduction goals into a single source of truth via the yearconfiguration table. This data drives UI logic (red...
Concurrent-users gauge + job-latency alerting after #38 #2529 2026-09-28 Abandoned 2026-09-28. Item A, a co2_active_users_5m gauge, is dropped: the manual scaling tiers it served are handled by the HPA and pgbouncer, people online is known by other means, and its reading depended on router affinity. Item B needs no change here: openshift-app-config#38 removed the job-class alerts, and plan prefill's requests belong in api.
Implementation Plan: Student FTE Calculator Helper #253 2025-12-17 Add calculator helper to student headcount form. Formula: (students × duration × avgfte) / 12 = Annual FTE. Button adds result to Total Student FTE input.
Implementation Plan: Improve Module Validation UX (Issue #265) #265 2026-01-05 Improve the module validation user experience by making the validation process more intuitive and moving key elements closer to where users are working.
Unified Implementation Plan: Generic Data Ingestion & Sync System #302 2026-01-20 This plan follows Spec-Driven Development: contracts first (DB, API, class interfaces), implementation second.
Complete Implementation Plan: API Travel Sync #302 2026-01-20 You need a plugin-based data ingestion system that can handle multiple data sources (CSV, Tableau API, future APIs) with a unified status tracking mechanism and background job processing.
Process Emissions Module — Implementation Plan #360 2026-02-19 Per issue #31 and the committee decision on 19-01-2026, the Infrastructure module is being split. The existing Infrastructure module (ID=3) continues to handle buildings. A new "Process Emissions" module (ID=8) is...
Implementation Plan: Reporting Aggregated Charts (#460) #460 2026-03-19 Both charts must react to the same filter state as the units table.
Percentage bar #460 2026-03-19 Implement a completion-percentage bar for Backoffice reporting using the existing completion-status pipeline, without introducing a new endpoint.
Implementation Plan: Reporting Aggregated Treemap Chart (#460) #460 2026-03-19 Add an Emission Breakdown treemap chart to the backoffice reporting page, visualising.
Implementation Plan: Issue #461 — Usage Statistics Box #461 2026-03-24 Then parse the result into {0: X, 1: Y, 2: Z}.
Treemap Charts Revert — Implementation Plan #498 2026-03-04 The endpoint response shape stays as defined in 252-chart-results-endpoint.md. No new keys are added.
Plan: PR #524 Review Fixes — Buildings Module v2 #524 2026-03-04 Address all review comments from guilbep on PR #524 (173-task-module-buildings-v2).
Emission Type Subcategory Granularity #577 2026-03-11 The current EmissionType hierarchy for Buildings is not granular enough. It needs a ZZ item level (room type) beneath each YY subcategory (energy type). Purchases also needs a new category.
Fix Travel Module After Factor Refactor (#621) #621 2026-03-12 Commit ddd7eb50 (refactor(factors): simplify factor seeding and handling, merged 2026-03-11).
Backoffice Merge Implementation Plan #630 2026-04-01 4. UI Texts Editing (new — receives Texts section from Documentation Editing).
Updated Charts Implementation Plan #632 2026-03-25 For modules with a single XX category (e.g., Equipment), the chart shows one bar with YY subcategories as segments.
Fix: Permission Conflict in Professional Travel (Issue #698) #698 2026-04-08 A "User Standard" with professionaltravel.edit permission is redirected to the.
Plan: 707 — Implement sync_module_factors (Backend only) #707 2026-04-08 Using example of Research facilities module, this feature aims at allowing a backoffice manager to update factors that are already stored in the database. The update logic is module data entry type dependent.
Plan: 707-ext — sync_module_factors for research_facilities common (DE type 70) #707 2026-04-08 Add ResearchFacilitiesCommonFactorUpdateProvider that recomputes kgco2eqsum.
Implementation Plan: Storybook Integration for Component Documentation #74 2026-01-04 Integrate Storybook 10.1.x into the Vue 3 + Quasar application to document and test components. This implementation will create a comprehensive component library similar to the EPFL Elements example, with full...
PRD 1 — Usage CSV (usage.csv) #760 2026-03-31 Provide a dataset that tracks the status and progress of each module for every carbon report.
Toggle Research Facilities #772 2026-04-07 Add a "Show Research Facilities" checkbox to the results page header (below "View Uncertainties") that shows/hides research facilities data in the carbon footprint charts and summary totals.
Health Check Endpoints - Quick Reference #773 2026-03-30 Then update backend/app/main.py to rename /ready back to /health.
Issue #773 - Health Check Refactor & Probe Logging Reduction #773 2026-03-30 Refactored health check endpoints across the entire application to reduce log noise and improve probe design according to PRD specifications.
Travel Bugs implementation plan #850 2026-04-15 This should be fixed and tested in all environments to make sure the behavior is consistent. Unit tests should be added to cover the different scenarios.
Issue #95: Deploy Documentation to Kubernetes #95 2026-04-01 Implement deployment of technical documentation (MkDocs) to Kubernetes alongside the main application, enabling branch-specific documentation accessible at /docs sub-path.
Autocomplete/Calculation for Professional Travel Destinations 2026-01-07 Add autocomplete functionality to the "from" and "to" fields in the professional travel form. The autocomplete shows airports when transportmode is "flight" and railway stations when transportmode is "train". Only...
Carbon Report Stats Aggregation Implementation 2026-03-19 Added a stats column to the CarbonReport model that aggregates emission statistics from all child CarbonReportModule records, similar to the existing module-level stats.
Carbon Report Stats and Progress Tracking Implementation 2026-03-19 4. Consistency: Report stats and progress automatically update when modules change.
Step 0 — Prep 2026-01-20 This gives you a clear separation of responsibilities.
detailed-view-reporting 2026-03-31 (see plan content)

Uncategorized (1)

Title Issue Last updated Summary
Backend-authoritative module activation for simulator report types #2494 2026-08-28 Explorer showed a greyed 0 bar for back-office-deactivated modules (#2465, fixed as a frontend-only patch in #2492). Investigation found the backend has no concept of back-office activation at all for is_simulator report types, and Planner has the identical bug, unmasked. Lays out three fix options and the two decisions (staleness, merge semantics) a maintainer needs to make before implementation.