Changelog

Strong beta branch notes followed by the main historical release log.

v0.4.130

2026-06-210.4.130 Release

featgeotech chat now auto-scans the current project folder at session start and seeds the agent with a bounded ground-model digest (strata values, lithology, groundwater, counts, verifier status), so the LLM understands your dataset from the first message; added --workspace/--no-workspace and /workspace + /rescan commands

featAdded --force-agent to geotech agent and geotech chat to skip the deterministic preflight short-circuit and route the request through the LLM agent loop; off by default

featRead-only data tools (query_ground_model, parse_ags/csv/cpt, read_file, analyze_signal_file) now serialize a wider ~8 KB result window and the hosted final-answer budget grows when a data tool was read, so the agent can reason across the whole dataset; data tools are paginated via limit/section

featgeotech analyze text output now signposts the AI agent verbs (geotech chat / geotech agent "<question>"); JSON and HTML output are unchanged

fixThe deterministic trust boundary is unchanged: the LLM interprets and orchestrates while deterministic code owns every number, and the new context/tool surfaces are read-only

v0.4.129

2026-06-210.4.129 Release

featAdded a read-only query_ground_model agent tool so the LLM agent can query the deterministic, evidence-bound GroundModel (strata, parameters, groundwater, SPT, coordinates, per-borehole detail) instead of re-parsing raw files

featEnriched the agent context with a bounded GroundModel values digest (representative strata, lithology, USCS, groundwater) plus a tool pointer, available to chat/agent and the swarm interpretation and reviewer roles

fixThe deterministic trust boundary is unchanged: the tool sources the model via analyzeWorkspace, never recomputes a number or writes a file, and the LLM interprets while deterministic code owns the values

v0.4.128

2026-06-140.4.128 Release

featPromoted the hosted strong-beta default text/agent model from glm-5.1 to glm-5.2, Z.ai’s latest flagship in the same GLM-5 series, served through the same chat-completions endpoint and hosted-beta proxy

fixUpdated the shared metadata default and supported-proxy-model catalog so the proxy allowlist, thinking-mode injection, CLI status/config output, and web defaults all follow glm-5.2; the superseded glm-5.1 is removed

fixRealigned the release-consistency guard, release smoke gate, BYOK/benchmark defaults, docs, and the active test corpus to glm-5.2; vision (glm-5v-turbo) and layout (glm-ocr) defaults are unchanged

v0.4.127

2026-06-140.4.127 Release

featAdded AGSi (AGS interchange ground-model JSON) and DIGGS 2.x (XML) export adapters that turn the normalized ground model into industry-standard geotechnical interchange files

featExposed both formats on the CLI (geotech export agsi / export diggs) and as agent tools (export_agsi / export_diggs), reusing the shared lithology vocabulary so geology codes, USCS, and unit colours stay consistent

fixFaithful-subset output by design: structurally correct and round-trippable, dependency-free, with full XSD validation, DIGGS measurements, and AGSi 3D geometry deferred

v0.4.126

2026-06-130.4.126 Release

featAdded a single deterministic lithology normalizer with a shared controlled vocabulary; the six previously independent soil/rock classifiers now delegate to it with unchanged outputs

fixFixed organic/peat classification: preserved as an organic class in the integrated review and recognized in the workspace dossier instead of being dropped to mixed/unknown

featPersisted normalized lithology on borehole layers and ground-model strata and surfaced it in borehole GeoJSON export, while always preserving the raw description

v0.4.125

2026-06-130.4.125 Release

featAdded deterministic page-break continuity repair for merged multi-page borehole logs with a typed audit trail, advisory findings, and unchanged blocking validation for genuine anomalies

featAdded deterministic coordinate plausibility validation covering axis-swap detection, BNG/UTM range envelopes, UTM zone vs longitude consistency, mixed-CRS detection, and cross-borehole spatial outlier flagging

fixExposed continuity repair thresholds through ingest options with conservative defaults and surfaced repair notes in the ingest dossier stratigraphy panel

v0.4.124

2026-06-050.4.124 Release

featAdded reviewed experimental excavation-plane-strain-dp-biot-replay drafting and backend execution for sequential Biot pressure-frame replay into Drucker-Prager effective stress

featAdded pressureReplayAudit result validation for accepted upstream Biot transient evidence, pressure audit metadata, zero nonlinear pore-pressure DOFs, and explicit limitations

fixExtended CLI aliases, agent schema normalization, default hosted GLM guardrails, route tests, and draft-run smoke coverage while FEM production readiness remains blocked

v0.4.123

2026-06-040.4.123 Release

featAdded an OpenGeoSys Liquid Flow h1_1Dsteady open-source comparison for steady Darcy seepage head, flux, and gradient evidence

featAdded a route-backed excavation-plane-strain-dp-adaptive draft path with hardening-modulus parsing and approval-bearing run templates

fixExtended default hosted GLM, agent normalization, CLI, routing, and draft-run smoke coverage for reviewed DP adaptive FEM execution

v0.4.122

2026-06-040.4.122 Release

fixMade persisted fem-reviewer-approval.v1 metadata mandatory for every reviewed experimental geotech fem run

fixUpdated FEM route, GroundModel workspace acceptance, docs, and agent command templates to include approval persistence metadata

fixExtended CLI and draft-run smoke coverage across raft, excavation, tunnel, consolidation, nonlinear-column, and Biot u-p scenarios

v0.4.121

2026-06-040.4.121 Release

featAdded check_fem_support_member_design as a deterministic agent-callable FEM support-member limit-state tool

featRequired explicit support demand provenance plus reviewer identity, license, jurisdiction, assumptions, and limitations before support capacities are computed

fixAdded default hosted GLM and swarm regressions so agents use deterministic support checks without claiming FEM production approval

v0.4.120

2026-06-040.4.120 Release

featAdded an OpenGeoSys staggered hydro-mechanics consolidation benchmark reference and open-source comparison record for Biot u-p pressure-profile evidence

featCompared the benchmark-scale Quad4 Biot pressure profile against the OpenGeoSys analytical p_D solution with deterministic hashes and series tolerances

fixExposed accepted external comparison IDs in FEM agent evidence summaries while keeping commercial-solver blockers active

v0.4.119

2026-06-040.4.119 Release

featForwarded reviewed hardeningModulusKpa values into the route-backed plane-strain Drucker-Prager adaptive backend

featAdded maxHardeningStressKpa result-envelope evidence and validation for hardened DP manifests

fixExtended core and CLI tests for persisted route-backed hardening evidence while keeping FEM productionReady false

v0.4.118

2026-06-040.4.118 Release

fixHardened persisted ingest job waiting so transient partial job.json reads during checkpoint-heavy large-PDF runs are retried below the CLI live-progress wrapper

fixAdded bounded persisted-read failure guidance with geotech ingest wait/resume instructions instead of surfacing raw JSON parser errors

featAdded isotropic hardening metadata to Drucker-Prager material-point, nonlinear-column, and plane-strain FEM evidence paths while full production FEM remains blocked

v0.4.117

2026-06-040.4.117 Release

fixAdded compact persisted ingest progress snapshots so large-PDF foreground polling can survive transient full job.json reads during checkpoint-heavy runs

featAdded an explicit load-generated-positive-pressure Biot envelope mode for mechanically loaded consolidation evidence while preserving the default pressure-envelope guard

featRegistered the load-generated Biot pressure acceptance benchmark in FEM evidence and default hosted GLM readiness prompts while full production FEM remains blocked

v0.4.116

2026-06-040.4.116 Release

featAdded per-step dense direct linear-solve residual audits to the plane-strain Biot u-p consolidation evidence kernel and transient acceptance metadata

fixHardened FEM reviewer approval validation for malformed validation summaries, duplicate assumptions or limitations, duplicate finding codes, and production-scope overclaims

fixFixed OpenAI-compatible BYOK support for localhost and private self-hosted endpoints while keeping remote endpoints fail-closed without an API key

v0.4.115

2026-06-040.4.115 Release

fixHardened borehole-log async ingest finalization so failed page checkpoints are recorded as review findings instead of replayed through the final merge

fixPreserved successful checkpoint data on large PDFs when a later visual page returns malformed provider JSON after retry

fixAdded regression coverage for mixed successful/failed borehole checkpoints with accurate successful and failed page counts

v0.4.114

2026-06-040.4.114 Release

featAdded generated published-source FEM benchmark comparison records for Terzaghi consolidation and alpha-zero Biot pressure dissipation with deterministic SHA-256 evidence/result hashes

featAdded sequential one-way Biot-pressure replay into the benchmark-scale Drucker-Prager effective-stress solve with explicit no-pressure-DOF audit metadata

fixAdded geotech fem run --require-approval-record and workspace acceptance metadata for strict reviewer approval persistence while full production FEM remains blocked

v0.4.112

2026-06-040.4.112 Release

fixHardened async PDF ingest finalization so completed page checkpoints replay from persisted OCR/text hints instead of making fresh OCR/model calls during final merge

fixCompacted per-page raw model diagnostic text stored in ingest job checkpoints to reduce large job.json churn on 100+ page borehole PDFs

fixAdded checkpoint-only borehole finalization regression coverage so large ingest jobs complete to reviewable results instead of surfacing transient JSON parse errors

v0.4.111

2026-06-040.4.111 Release

featAdded reviewed experimental geotech fem run --backend plane-strain-dp-adaptive for explicit 2D Drucker-Prager excavation cases with adaptive cutback rollback audit

featHardened result validation for adaptive DP metadata, solver residuals, full-load acceptance, rollback signatures, and productionReady overclaims

fixExtended FEM readiness and default hosted GLM guardrails so nonlinear evidence is visible while full production FEM remains blocked

v0.4.110

2026-06-040.4.110 Release

featPromoted Biot transient acceptance metadata into result manifests, validation, CLI output, draft/run smoke, and WebGL smoke coverage

featAdded deterministic support member yield, buckling, flexure, slenderness, and combined utilization evidence with reviewer/license metadata validation

fixDowngraded repeated malformed provider JSON on long visual PDF ingest pages to manual review after retry instead of failing the whole async job

v0.4.109

2026-06-040.4.109 Release

featAdded committed Gauss-point Drucker-Prager state history with strain increments and accumulated plastic-strain evidence

featAdded non-monotonic loadHistoryFactors coverage for unload/reload FEM evidence without resetting plastic state to virgin material

fixUpdated FEM readiness summaries so path-dependent plasticity evidence is visible while full production FEM remains blocked

v0.4.108

2026-06-030.4.108 Release

featAdded audited CSR sparse linear-algebra utilities and an opt-in sparse Conjugate Gradient backend for benchmark-scale Drucker-Prager plane-strain load steps

featAdded structured external benchmark acceptance metadata with published/commercial source requirements, quantity tolerances, and readiness blockers

fixStrengthened staged excavation support screening and mocked default hosted GLM guardrails while keeping full production FEM blocked

v0.4.107

2026-06-030.4.107 Release

featAdded explicit nonlinear FEM convergence reporting with per-load-step residual histories, termination reasons, and fail-closed nonconvergence metadata

featHardened the experimental Biot u-p preview with bounded transient step policy, pore-pressure diagnostics, pressure-overshoot rejection, and manifest validation

fixBlocked production-design FEM approval scope in reviewer records and CLI runs until production solver, benchmark, support-design, and jurisdiction-specific gates are implemented

v0.4.106

2026-06-030.4.106 Release

featAdded route-backed experimental geotech fem demo biot and geotech fem run --backend biot-up paths for reviewed plane-strain Biot u-p seepage/groundwater cases

featUpdated GroundModel, ingest dossier, benchmark, production-readiness, and mocked default GLM surfaces so seepage/groundwater coupling appears as a human-reviewed preview route

fixHardened large foreground ingest jobs by extending bounded transient job.json read retries and reusing the last valid live-progress snapshot through partial checkpoint reads

v0.4.105

2026-06-030.4.105 Release

featAdded an alpha-zero Quad4 Biot u-p drainage-column benchmark that checks transient pore-pressure dissipation against Terzaghi average consolidation at Tv = 0.197

featRegistered the pressure-dissipation benchmark in FEM engineering evidence and production-readiness surfaces so hosted GLM and BYOK-style agents can see the stronger consolidation evidence

fixKept full production FEM blocked: the Biot kernel remains benchmark-scale evidence, not a route-backed production sparse solver, nonlinear staged-construction backend, support-design engine, or independently approved benchmark corpus

v0.4.104

2026-06-030.4.104 Release

featHardened the benchmark-scale Quad4 Biot u-p evidence kernel with explicit excess-pore-pressure, stress, storage, and Darcy-flux convention metadata plus auditable pressure-equation residual components

featAdded deterministic Biot pressure-gradient flux and alpha-zero decoupling benchmark coverage and a fail-fast unsupported negative free-pore-pressure policy

fixExtended production-readiness and mocked default GLM/BYOK agent tests so stronger Biot evidence is visible while full production FEM remains blocked behind route, sparse-solver, nonlinear-coupling, benchmark, support-design, and approval gates

v0.4.103

2026-06-030.4.103 Release

fixAdded foreground geotech ingest <pdf> regression coverage for transient partial job.json reads so the initial command path keeps polling instead of aborting with a raw Unterminated string in JSON error

fixConfirmed the reported ingest failure was from an older global geotech 0.4.99 binary; the published latest CLI includes the live-progress hardening

fixKept FEM production readiness unchanged while the next solver work remains evidence-hardening and validation, not a production claim

v0.4.102

2026-06-030.4.102 Release

featAdded a benchmark-scale Quad4 plane-strain Biot u-p backward-Euler evidence kernel with coupled displacement and pore-pressure DOFs, pressure storage, Darcy pressure-gradient flow, and Biot stress reduction

featExtended FEM evidence and mocked default GLM planning summaries with Biot u-p benchmark IDs, residual/mass-balance metrics, and stress sign-convention coverage

fixKept full production FEM blocked until route-backed manifests, production sparse solvers, nonlinear plasticity coupling, independent benchmarks, support design, and enforced approvals are complete

v0.4.101

2026-06-030.4.101 Release

fixFixed large resumable PDF ingest live progress so transient partial job.json reads are skipped and retried instead of aborting with a raw Unterminated string in JSON error

fixAdded regression coverage for foreground ingest wait recovery while the background worker continues to a completed persisted result

fixKept repeated unreadable persisted job state bounded with resume/wait guidance rather than silently hanging

v0.4.100

2026-06-030.4.100 Release

featAdded a benchmark-scale Quad4 plane-strain steady Darcy seepage evidence kernel with hydraulic head DOFs, Darcy flux, pore pressure, and effective-stress-reduction metadata

featExtended FEM evidence with 2D linear-head flow, boundary mass-balance, and effective-stress-reduction benchmarks visible to readiness tools and default hosted GLM planning

fixKept full production FEM blocked until route-backed result manifests, Biot u-p coupling, sparse production solvers, external benchmarks, support design, and enforced approvals are complete

v0.4.99

2026-06-030.4.99 Release

featAdded a benchmark-scale mechanical-only Quad4 plane-strain Drucker-Prager load-step kernel with Gauss-point plastic state, residual/reaction checks, and collapse detection

featExtended FEM evidence with elastic-regression, shear plastic-patch, global residual, staged plastic-strain monotonicity, and collapse-detection benchmarks

fixKept full production FEM blocked until sparse solver routes, Biot/pore-pressure coupling, consistent tangent approval, external benchmarks, support design, and enforced approvals are complete

v0.4.98

2026-06-030.4.98 Release

featAdded a benchmark-scale linear Quad4 plane-strain global assembly kernel with affine patch, skewed geometry, loaded reaction-balance, free-residual, dense-size-cap, and fail-fast validation coverage

featWired global assembly evidence into FEM engineering evidence, production-readiness summaries, exports, and mocked default GLM planning tests

fixKept full production FEM blocked until nonlinear global solver coupling, pore-pressure/consolidation coupling, production sparse solver/result routes, support-design standards, persisted approvals, and independent benchmarks are approved

v0.4.97

2026-06-030.4.97 Release

fixHardened hosted-beta, Z.ai, OpenAI-compatible, Anthropic, and Hugging Face adapters so truncated provider JSON is reported as a provider response failure instead of a raw parser error

fixMade async PDF ingest retry malformed provider JSON once at the page level before recording a resumable page failure

fixAdded regression coverage for malformed hosted/GLM responses and the reported Unterminated string in JSON failure mode

v0.4.96

2026-06-030.4.96 Release

featAdded a Drucker-Prager/Mohr-Coulomb-compatible nonlinear material-point return-mapping kernel with principal stress, plastic strain, and yield-residual metadata

featAdded geotech fem run --backend nonlinear-column for reviewed staged consolidation cases with deterministic solver load-step, iteration, force-residual, yield-residual, and plastic-strain metadata

fixKept FEM production readiness blocked until global 2D/3D assembly, pore-pressure DOFs, independent benchmarks, support design, and enforced production approvals are complete

v0.4.95

2026-06-030.4.95 Release

featAdded an experimental staged settlement/consolidation FEM preview route with deterministic 1D Terzaghi time stepping, staged load history, settlement envelopes, and pore-pressure metadata

featAdded geotech fem demo consolidation plus staged consolidation draft/run and WebGL/fallback smoke coverage while keeping full 2D/3D coupled production FEM blocked

fixUpdated GroundModel, ingest dossier, benchmark, and default hosted GLM agent paths so staged consolidation is treated as a human-reviewed preview route, not a contract-only or production solver route

v0.4.94

2026-06-030.4.94 Release

fixFixed resumable ingest job checkpoint persistence so large job.json writes stay atomic while the CLI polls live progress

fixAdded transient persisted-job JSON read retries and job-aware resume guidance instead of surfacing bare parser errors on partial checkpoint reads

fixHardened project storage with the same atomic-write and transient-read policy for project-backed ingest reviews, artifacts, notes, and active context

v0.4.93

2026-06-030.4.93 Release

featAdded deterministic FEM engineering evidence kernels for Mohr-Coulomb material-point plasticity, 1D Terzaghi consolidation, 1D Darcy seepage, effective-stress coupling, support screening, tolerance policy, and reviewer approval validation

featAdded persisted FEM approval-record support so reviewed preview runs can carry reviewer identity, license, jurisdiction, assumptions, limitations, validation summary, and case hash

fixKept FEM production readiness blocked until the verified kernels are coupled to solver routes and independently benchmarked against published or commercial references

v0.4.92

2026-06-030.4.92 Release

featAdded a deterministic FEM production-readiness assessment for nonlinear/plasticity, consolidation, seepage/pore-pressure coupling, advanced staging, support design, real project workspace-to-run, benchmark, and reviewer gates

fixKept full production-grade FEM claims blocked until solver physics, independent benchmarks, reviewed workspace acceptance, and licensed-review records are implemented

fixRequired geotech fem run to include --experimental --reviewed and extended FEM smoke coverage with contract-only route checks, reference balances, and cross-scenario trend assertions

v0.4.91

2026-06-030.4.91 Release

featAdded production-candidate guardrail contracts for preprocessing, corpus benchmarking, BYOK, standards profiles, calculation drafts, planned FEM routes, and deterministic signal analysis

fixAdded runtime OpenRouter and Hugging Face BYOK aliases so users can bring OPENROUTER_API_KEY, OPENROUTER_MODEL, and HUGGINGFACE_API_KEY into real CLI sessions

fixKept AI-assisted and FEM-derived outputs review-gated, path-safe, and free of raw prompt, response, model payload, private path, and token-shaped leakage in persisted artifacts

v0.4.90

2026-06-010.4.90 Release

featAdded a provider-neutral BYOK benchmark contract using the same preprocessed page-evidence prompt across hosted-beta, OpenAI-compatible, OpenRouter/free, and local/HF-compatible routes

fixMade BYOK response validation require compact JSON, source-evidence citations, preserved review gates, and no claims of direct image/PDF inspection

featExtended npm run smoke:byok with structured JSON report output for provider profile, model, latency, token, and evidence-contract comparison

v0.4.89

2026-06-010.4.89 Release

fixAdded a machine-readable path-safety block to the internal geotechnical corpus benchmark report

fixMade corpus acceptance fail closed when fixture or benchmark inputs contain local absolute paths or secret-shaped values

fixKept path and secret findings redacted in JSON and HTML output so private fixture paths and BYOK tokens are not echoed into persisted artifacts

v0.4.88

2026-06-010.4.88 Release

featSplit the internal geotechnical corpus benchmark provider matrix into hosted-beta, OpenAI-compatible BYOK, OpenRouter/free, and local/HF-compatible profiles

fixAdded fail-closed provider evidence checks so text-only, free, and local profiles must consume preprocessed OCR/page evidence instead of direct image or native-PDF tasks

featAdded corpus-level confidence component summaries and trend fields for extraction, traceability, corroboration, readiness, and page-evidence trust review

v0.4.87

2026-06-010.4.87 Release

fixFixed the trusted-publishing release build by giving the web package a temporary npm lockfile during Next.js build so the Node 24 publish job selects npm for SWC registry fallback

fixKept the public web build behavior unchanged while aligning the release pipeline with the already-passing CI build

v0.4.86

2026-06-010.4.86 Release

featAdded path-safe signal benchmark history and trend artifacts for deterministic monitoring benchmark comparisons over time

featAdded a signal-trend HTML dashboard with instrument coverage and run-to-run deltas beside the JSON and SVG benchmark outputs

fixKept signal benchmark acceptance fail-closed on path leaks while preserving zero-model-call checks for all five synthetic instrument classes

v0.4.85

2026-05-310.4.85 Release

featAdded opt-in deterministic signal threshold profiles for settlement, piezometer, inclinometer, vibration PPV, and load-test monitoring data

featExtended geotech signal analyze and the sandboxed analyze_signal_file tool with threshold-profile auto and named review profiles while keeping signal metrics deterministic

fixExpanded the signal benchmark fixture to all five instrument classes and made comparison/direct-signal benchmark artifacts path-safe

v0.4.84

2026-05-310.4.84 Release

featExtended opt-in GroundModel calculation input drafts with source references, grouped source pages, evidence-derived confidence, and explicit review gates

fixKept calculation and FEM drafts non-executing until missing user inputs, assumptions, and evidence gates are resolved

fixIncreased the broader preprocessing v2 fixture regression timeout so full core CI has enough budget for deterministic multi-PDF region rendering

v0.4.83

2026-05-310.4.83 Release

featAdded deterministic standards-profile validation for Eurocode 7, AASHTO, IS, BS, and ASTM readiness with blocker codes, required assumptions, safety-factor context, and source references

featWired standards-profile validation into GroundModel verification while keeping generated calculation drafts review-gated and non-executing

fixAdded focused regression coverage for profile assumptions, missing evidence blockers, workflow blockers, and source-reference propagation

v0.4.82

2026-05-310.4.82 Release

featAdded deterministic mixed digital/scanned and malformed scanned PDF fixtures so preprocessing v2 coverage spans borehole, CPT, lab, mixed, and malformed report page shapes

featAdded npm run benchmark:preprocessing to compare none, ocr-optimized, and region-v2 without hosted model calls, writing JSON, HTML, and SVG summaries

fixMade preprocessing benchmark output path-safe by storing fixture basenames and hashes only and failing if local paths or secrets leak into persisted output

v0.4.81

2026-05-310.4.81 Release

featAdded deterministic scanned CPT table, lab table, and mixed scanned-report PDF fixtures to broaden region-v2 crop-detection coverage

fixAdded preprocessing-only regression coverage proving region-v2 detects normalized crop assets and region quality across broader geotechnical scanned layouts without hosted model calls

fixUpdated development status so preprocessing v2 broader-fixture coverage remains visible as the next hardening slice

v0.4.80

2026-05-310.4.80 Release

featAdded a deterministic in-repo image-only borehole/table PDF fixture so region-v2 preprocessing is tested against rendered PDF evidence, not only synthetic benchmark metadata

fixMade corpus acceptance fail when region-v2 returns zero preprocessing regions, zero persisted crop assets, low region quality, or traceability below none and ocr-optimized

fixUpdated corpus HTML comparisons with region quality deltas and kept committed public fixture outputs free of local absolute path leaks

v0.4.79

2026-05-310.4.79 Release

featAdded explicit region-v2 PDF/image preprocessing with fine-pass deskew metadata, higher-resolution normalized pages, smarter table and borehole/log strip crop candidates, and region quality scoring

featRoutes region-v2 table/log crop assets into OCR before full-page vision OCR so borehole and table evidence gets a tighter first extraction target while retaining provider-neutral fallback behavior

fixExtended corpus benchmarks to compare none, ocr-optimized, and region-v2 by default across region quality, persisted crop assets, deskew, hosted-call, traceability, and GroundModel readiness deltas

v0.4.78

2026-05-310.4.78 Release

featAdded npm run benchmark:geotech-corpus as an internal R&D wrapper over the geotechnical document benchmark contract across fixture categories, provider profiles, preprocessing modes, traceability, cache reuse, GroundModel readiness, and FEM execution-boundary guardrails

featAdded real-fixture corpus mode for private local PDFs and images supplied through GEOTECHCLI_BENCHMARK_* env vars, including first-pass plus cached-rerun execution and none versus ocr-optimized preprocessing comparison

fixKept private report bytes and local absolute paths out of persisted corpus benchmark JSON while still writing local history, trend, SVG, and HTML summaries for repeatable review

fixPromoted review gates into corpus run summaries so benchmark acceptance captures trust boundaries as well as latency, hosted calls, cache hits, traceability, GroundModel readiness, and region quality

v0.4.77

2026-05-250.4.77 Release

fixKept all recovered report boreholes selectable across integrated ingest evidence, strip-log, extracted-field, map, and A-A section views instead of collapsing to the first structured borehole

fixPromoted borehole-specific coordinate text from retained OCR/native report evidence into GroundModel map points, including compact latitude/longitude OCR rows and projected easting/northing rows

fixKept site-only coordinates out of borehole maps and preserved schematic alignment fallback when recovered borehole coordinates are missing or partial

featAdded geotech agent --task calculation-readiness and prompted routing for bearing, settlement, pile, liquefaction, slope, and FEM draft readiness over deterministic GroundModel verifier evidence

v0.4.75

2026-05-240.4.75 Release

fixAdded regression coverage for failed geotech agent --route-with-model router calls so outage paths still fall back to the workspace-backed agent and keep a failed router row in model_calls.jsonl

fixClarified project-aware agent guidance so fallback provenance describes the current planned workspace-backed handoff row without implying unavailable post-run model-call telemetry

v0.4.74

2026-05-240.4.74 Release

featAdded geotech agent --route-with-model so ambiguous project prompts can request a strict JSON workflow route proposal from the configured hosted or BYOK model

featWorkflow route artifacts now record selectionSource, rejected model tasks, confidence-gate status, and opt-in router proposal rows in model_calls.jsonl

fixRejected unknown model-supplied workflow tasks, kept no-evidence model proposals below the deterministic execution gate, and preserved fallback to the workspace-backed agent path

v0.4.73

2026-05-240.4.73 Release

featAdded a provider-neutral project workflow router so prompted project requests can select and sequence deterministic workflows before any optional model review

featPrompted requests such as geotech agent "find anomalies and create visualizations" now write workflow_route.json, deterministic workflow outputs, and a combined workflow_route_report.md

fixAdded strict route validation and provider-contract tests so hosted GLM, BYOK OpenAI-compatible routes, direct Z.ai, and local-compatible models share the same allowed workflow task contract

v0.4.72

2026-05-240.4.72 Release

featAdded a provider-neutral project workflow executor for explicit geotech agent --task data-quality, ground-model, risk-analysis, anomaly-detection, recommendations, and visualization runs

featAdded deterministic .geotech workflow_result.json, workflow_report.md, and workflow_trace.json artifacts while keeping model_calls.jsonl empty for task runs

fixAdded core project-workflow exports and regression coverage for deterministic workflow results, chart specs, reports, and CLI short-circuit behavior

v0.4.71

2026-05-230.4.71 Release

fixAdded explicit repository read permissions to release verify and Cloudflare deploy jobs so checkout works when default workflow permissions are tightened

fixClarified geotech agent --no-workspace help text so it matches the current opt-out behavior for automatic project-aware discovery

v0.4.70

2026-05-230.4.70 Release

featAdded project-agent root detection across explicit workspace, .geotech/project.json, git root, and cwd so project-aware runs have a clear filesystem boundary

featMade prompted geotech agent runs project-aware by default unless --no-workspace is supplied, with geotech agent . entering discovery mode

featAdded project-agent intent artifacts, run manifests, file/evidence JSONL indexes, local memory.json, tool-call traces, and model-call traces under .geotech

fixStopped swarm plans from advertising disabled skill tools or unavailable orchestrator deliverable tools, preserved reviewer tool context, and kept rejected reviews unresolved instead of approved-with-notes

v0.4.69

2026-05-230.4.69 Release

featAdded project-aware geotech agent discovery: --plan-only and no-prompt runs scan the workspace, write .geotech project state, and show workflow readiness before any LLM call

fixUpdated FEM draft recommendations so reviewed cases point to geotech fem run <analysis_case.json> --experimental instead of stale built-in demo commands

fixAligned OpenAI-compatible BYOK config with docs by using OPENAI_COMPATIBLE_MODEL as primary and OPENAI_COMPATIBLE_MODEL_ID as a fallback alias

v0.4.68

2026-05-210.4.68 Release

fixFixed the sequential npm publish script so each workspace publishes from its package directory instead of an ambiguous package spec

fixNormalized published package repository metadata with explicit GitHub URLs and workspace directories for @geotechcli/core and geotechcli

securityKept the tokenless Trusted Publishing path and moved the repair release to 0.4.68 after 0.4.67 failed before publishing npm packages

v0.4.67

2026-05-210.4.67 Release

securityReplaced concurrent Changesets publishing with a deterministic npm publish script that publishes @geotechcli/core first and waits for registry visibility before publishing geotechcli

fixRestricted npm publishing to the strong-beta branch so tag-triggered workflows create GitHub releases without attempting a second npm publish

fixAdded package-version and dependency guardrails so the CLI cannot publish against a missing or mismatched core package version

fixMade the Cloudflare beta deploy wait for successful npm publishing so the live site cannot advance after a partial npm release

v0.4.66

2026-05-210.4.66 Release

fixFixed the Trusted Publishing toolchain guard quoting so Bash does not expand JavaScript template expressions before Node evaluates the release check

securityKept the npm publish path tokenless and OIDC-based for the second Trusted Publishing release attempt

v0.4.65

2026-05-210.4.65 Release

securitySwitched the npm publish job to Trusted Publishing/OIDC by removing the long-lived NPM_TOKEN publish path from changeset publish

securityAdded a Node 24/npm 11.5.1+ release toolchain check so OIDC publishing fails early if the runner cannot satisfy npm Trusted Publishing requirements

fixAdded GitHub repository metadata to the published core package so npm can match the OIDC trusted publisher to the repository

docsDocumented the exact npm Trusted Publisher settings for both published packages before the old write token is revoked

v0.4.64

2026-05-190.4.64 Release

featAdded GroundModel-derived FEM draft candidates to report and borehole ingest output for foundation settlement and staged excavation deformation routes

featAdded a FEM draft routing table to HTML report output with readiness, missing user inputs, review gates, evidence prefill, and draft commands

fixKept LLM and swarm boundaries unchanged: agents can plan, draft, and validate FEM cases, while deterministic execution stays behind human-invoked geotech fem run --experimental

v0.4.63

2026-05-190.4.63 Release

featAdded geotech fem run <analysis_case.json> --experimental so reviewed FEM draft cases can execute through deterministic built-in preview backends

featAdded a draft-to-run acceptance smoke harness covering raft, excavation, and tunnel case files, deterministic manifests, and HTML artifact output

fixKept LLM and swarm FEM authority bounded to planning, drafting, and validation; agents can recommend a reviewed run but cannot execute solver/WebGL generation as a tool

docsUpdated FEM docs and examples to distinguish draft, reviewed run, and built-in demo workflows while preserving the non-design safety boundary

v0.4.62

2026-05-180.4.62 Release

featAdded geotech fem demo tunnel --experimental as the third deterministic 3D preview, generating an empirical tunnel volume-loss settlement surface without hosted model calls

featAdded a tunnel volume-loss analysis-case contract, deterministic result manifest, finite visualization mesh, and validation gates for diameter, axis depth, alignment length, volume loss, and trough-width factor

featExtended geotech fem draft tunnel-volume-loss-settlement so agents and users can prepare review-gated tunnel settlement drafts from explicit inputs

fixUpdated the FEM WebGL and Canvas smoke harness to generate and test raft, excavation, and tunnel artifacts at desktop and mobile sizes

fixKept the tunnel preview empirical and non-design: LLMs may route, draft, and review the case, but geotechCLI deterministic code owns the generated settlement surface

v0.4.61

2026-05-180.4.61 Release

featAdded deterministic simulation-tool context to reviewer swarm prompts so FEM drafts and validation data remain visible even when a model handoff summary is terse

featAdded compact per-tool summaries for FEM and calculation outputs before reviewer JSON context, keeping BYOK and hosted models aligned on the same evidence contract

fixAdded runtime swarm regressions proving simulation agents can prepare FEM cases, reviewer agents can validate FEM cases, and reviewer agents cannot prepare FEM cases

fixKept FEM safety boundaries unchanged: LLMs plan and review while geotechCLI deterministic contracts own case drafting, validation, and future solver outputs

v0.4.60

2026-05-180.4.60 Release

featAdded a GroundModel-to-FEM draft adapter that converts FEM readiness drafts into provider-neutral prepare_fem_analysis_case inputs with retained evidence references

featAdded geotech fem draft <objective> --workspace <dir> so workspace evidence can prefill material, groundwater, and source traceability before users provide geometry, loads, and staging

featAdded geotech fem agent ... --workspace <dir> to give the scoped FEM brain FEM-only GroundModel readiness context without widening its tool allowlist

fixKept FEM safety boundaries unchanged: no solver auto-run, no design approval, and no LLM-invented displacement, reaction, mesh, or stage values

v0.4.59

2026-05-170.4.59 Release

featAdded geotech fem draft <objective> for deterministic review-gated FEM analysis-case drafts without running a solver or creating WebGL results

fixWired FEM GroundModel readiness command templates to draft analysis_case.json inputs before experimental demos

fixHardened FEM agent validation so blocked cases return inspectable reviewer data instead of tool transport failures

featAdded compact FEM draft and validation summaries for LLM and swarm handoff while keeping FEM math deterministic

v0.4.58

2026-05-170.4.58 Release

featAdded optional resultFields, steps, and datasets metadata to FEM result manifests so future FEM previews can share one viewer and validation contract

fixKept the existing visualization disp/color/frames contract backward-compatible while validating malformed result metadata when present

featAdded geotech fem agent as a scoped LLM FEM planning surface that can list FEM capabilities, prepare analysis-case drafts, and validate cases only

featAdded scripts/smoke-fem-webgl.mjs and npm run smoke:fem:webgl for reusable desktop/mobile FEM artifact QA

v0.4.57

2026-05-170.4.57 Release

featAdded geotech fem demo excavation --experimental as the second deterministic 3D FEM/WebGL preview for staged excavation deformation screening

featAdded an excavation analysis-case contract, staged elastic demo runner, finite-result validation, and manifest fields for surface settlement, horizontal displacement, wall-deflection proxy, support reaction, and stage count

featExtended the FEM WebGL artifact with optional field and stage controls for staged manifests while preserving the existing raft preview layout

fixUpdated FEM routing so agents and swarm roles can prepare review-gated excavation-deformation drafts without inventing FEM result values

fixAdded GroundModel readiness and non-executing input drafts for experimental staged excavation FEM so workspace analysis no longer exposes FEM as foundation-only

v0.4.56

2026-05-170.4.56 Release

featAdded provider-neutral FEM capability routing so agents can list implemented, draft, and planned FEM routes before attempting numerical work

featAdded list_fem_capabilities, prepare_fem_analysis_case, and validate_fem_analysis_case agent tools for deterministic FEM planning, case drafting, and review-gated validation

featWired FEM planning into the simulation swarm allowlist and FEM validation into the reviewer allowlist so LLMs route and review while geotechCLI owns contracts and validators

featAdded an experimental fem-foundation-settlement GroundModel readiness workflow with non-executing input drafts and review gates

fixFixed swarm planner readiness extraction for ready_with_assumptions workflows and threaded FEM route/validation tools into DesignEngineer and RiskReviewer plans

v0.4.55

2026-05-160.4.55 Release

featAdded geotech fem demo raft --experimental as an opt-in deterministic 3D FEM/WebGL preview for raft settlement workflow development

featAdded a provider-neutral FEM analysis-case and result-manifest contract in core with validation for assumptions, finite result arrays, mesh references, and reaction-balance review

featAdded a self-contained WebGL exporter for the experimental raft preview with deformation scale, mesh wireframe, raft load patch, validation warnings, assumptions, and limitations

fixDocumented the FEM preview as experimental non-design behavior in README and website docs while keeping production calculations and report workflows unchanged

fixAdded regression coverage for the deterministic raft manifest, WebGL export contract, CLI experimental gate, JSON output, and artifact writing

v0.4.54

2026-05-150.4.54 Release

featReplaced the legacy dark ingest dossier with the light integrated geotechnical review interface for HTML ingest output

featPersisted compact GLM-OCR layout pages through geotechnical document ingest, borehole ingest, page audits, and page evidence cache reuse

featConverted persisted OCR layout geometry into source-page overlays automatically with conservative text-matched evidence linking for boreholes, extracted parameters, and image/table regions

fixHardened source-region rendering for duplicate GLM indices, stale region links, invalid persisted pages, ratio bboxes, and GLM layout priority over older reconstructed evidence boxes

fixNarrowed hosted-beta web proxy imports to @geotechcli/core/llm, removing the Next.js critical dependency warning from the web production build without changing CLI PDF or vision runtime loading

securityKept the release on stable Next.js after audit review: the high and critical npm audit gate is clean, while the remaining moderate Next/PostCSS advisory is a tracked upstream nested dependency with no stable patched Next release available yet

fixAdded regression coverage for cache layout round-tripping, geotechnical document cache hits, borehole page audits, integrated layout source pages, and Playwright-verified desktop and mobile review output

v0.4.53

2026-05-150.4.53 Release

securityMade geotech agent and geotech chat treat --skills as the authoritative per-session gate for installed skill tools

fixKept direct geotech skill commands live while preventing persisted skill runtime settings from leaking into LLMConfig skill access

fixAdded CLI regression coverage for agent, chat, and direct skill command behavior around skill opt-in and approved-skill execution

fixReplaced several Windows-hostile CLI glyphs in agent and vision status output with ASCII text and fixed a corrupted development-status apostrophe

securityUpgraded the web workspace to the May 2026 Next.js security patch line and pinned patched PostCSS/AWS XML parser dependencies used by the website build chain

v0.4.52

2026-05-100.4.52 Release

featAdded provider-neutral Confidence v2 for geotechnical document ingest results with extraction, page evidence, traceability, corroboration, engineering completeness, readiness, missing-data, and review-gate scores

featThreaded confidenceBreakdown into DocumentEvidencePacket, compact agent summaries, benchmark JSON, CLI summaries, and generated HTML reports

featRenamed report confidence presentation to Review confidence and added a compact Trust breakdown section before GroundModel visuals

fixTightened missing-critical-data scoring so not-reported values do not count as usable engineering evidence

fixUpdated the canonical GeotechnicalInvestigationReport cached benchmark fixture with Confidence v2 data and verified the real PDF benchmark with 34/34 pages, 100% cache hits, zero hosted calls, and 100% direct source-page traceability

fixAdded Playwright as a development dependency and verified the generated Confidence v2 HTML report at desktop and mobile widths with no page-level horizontal overflow

v0.4.51

2026-05-090.4.51 Release

featAdded a report-side GroundModel adapter for geotechnical PDF ingest results with source-page evidence IDs, inferred strata, SPT N-values, depth-bound parameters, and groundwater observations

featAdded a GroundModel Visual Review section to generated geotechnical report HTML with compact borehole strip logs, SPT N-value depth plotting, lab-parameter depth charts, and groundwater summaries

fixKept the DocumentEvidencePacket schema stable while wiring richer visual review through the report view model

fixAdded regression coverage for report HTML GroundModel visual rendering from PDF evidence

v0.4.50

2026-05-090.4.50 Release

featAdded a GroundModel Visual Review section to geotech analyze --format html with compact borehole strip logs, SPT N-value depth plotting, lab-parameter depth mini charts, and groundwater/monitoring summaries

featExtended geotech viz GroundModel support beyond coordinate maps so analyze JSON now renders map, SPT-depth, lab-depth, and groundwater charts from the same provider-neutral GroundModel contract

fixAdded context-aware table inference so common coordinate headers such as id and groundwater files with generic depth_m columns still bind to borehole map points and groundwater observations

fixKept visual outputs evidence-first by carrying borehole IDs, depth, confidence, warnings, and evidence IDs into SVG titles and interactive chart metadata

fixAdded regression coverage for richer HTML visual review and multi-chart GroundModel visualization output

v0.4.49

2026-05-090.4.49 Release

featAdded a provider-neutral ground-model-map.v1 contract with borehole coordinate points, CRS/local-grid warnings, confidence, source evidence references, and plottable extents

featAttached GroundModel map data during geotech analyze so workspace analysis carries spatial evidence alongside boreholes, standards readiness, and calculation-routing context

featAdded a GroundModel Map section to geotech analyze --format html with plan-view SVG coordinates, map metrics, warnings, and traceable coordinate rows

featUpdated geotech viz so GroundModel JSON and analyze JSON are detected automatically and rendered as an interactive coordinate map with point labels and evidence metadata

fixAdded regression coverage for GroundModel map creation, analyze HTML rendering, and map chart generation

v0.4.48

2026-05-060.4.48 Release

featAdded a deterministic role-based swarm execution plan for geotech agent --swarm with WorkspaceScout, DataEngineer, GroundModeler, StandardsChecker, DesignEngineer, RiskReviewer, and ReportEngineer ownership

featThreaded workspace evidence, standards profile, calculation readiness, missing inputs, and approved executable skills into swarm prompts before specialist execution starts

featMade geotech agent --workspace ... --swarm request non-executing calculation input drafts during workspace analysis so ready and blocked workflows are explicit

fixExcluded prompt-only and unapproved skills from swarm execution while preserving them as audit warnings

fixAdded regression coverage for swarm plan prompt injection, role ownership, skill selection, and workspace readiness formatting

v0.4.47

2026-05-060.4.47 Release

featAdded standards-profile assumptions for Eurocode 7, AASHTO, IS, BS, and ASTM workspace analysis so GroundModel readiness carries an explicit design-profile basis

featAdded opt-in, non-executing calculation input drafts for bearing, settlement, pile, liquefaction, and slope workflows through geotech analyze --draft-inputs

fixRepaired bundled skill bootstrap so partial installs no longer stop after the first few skill manifests and first use repairs the complete approved catalog

fixRemoved the stale public docs example that referenced an internal bundled ZIP filename

fixAdded regression coverage for standards profile exports, calculation input drafts, workspace propagation, HTML readiness rendering, and partial skill-catalog repair

v0.4.46

2026-05-050.4.46 Release

featAdded a provider-neutral DocumentEvidencePacket synthesis prompt compiler so report synthesis reads the same page, observation, source-page, review-gate, and traceability contract used by agents and benchmarks

fixRouted geotechnical document synthesis through the evidence-packet compiler instead of raw extraction arrays, preserving whole-report outline order, borehole continuity, missing-data gates, and source-page citation rules across hosted GLM and future BYOK providers

featAdvanced the document evidence packet schema to v2 with extracted risk and recommendation signals so synthesis can stay packet-first without losing engineering interpretation context

featAdded a provider-agnostic agent operating contract for single-agent, swarm, and legacy specialist prompts so BYOK models receive the same GeotechCLI evidence, tool, capability, confidence, and review-gate instructions as hosted GLM

fixAdded free/open-route model adaptation rules for compact evidence, smaller steps, provider capability failures, and native-PDF/image fallbacks so free OpenRouter-style models fail clearly instead of drifting silently

fixTightened OpenAI-compatible capability profiling so known text-only free routes are review-gated for image understanding while multimodal omni/VL/vision routes remain image-capable

featAdded npm run smoke:byok for local provider-key smoke tests across Z.ai, OpenAI, Anthropic, Hugging Face, and OpenAI-compatible endpoints when matching environment keys are present

fixDocumented BYOK smoke environment variables and added regression coverage for packet-first synthesis prompt compilation and provider-agnostic agent prompt injection

v0.4.45

2026-05-050.4.45 Release

featAdded a schema-validated DocumentEvidencePacket for geotechnical document ingest results, normalizing page methods, source pages, observations, content chunks, synthesis, review gates, and traceability into a provider-neutral contract

fixAttached evidence packets to geotechnical document ingest and async job-adjusted results so reports, benchmarks, and agents share the same evidence shape

featAdded compact agent evidence summaries for geotechnical ingest, job-result, and persisted-review tools so source pages, methods, missing values, review gates, borehole IDs, and max depth survive prompt truncation

fixUpdated single-agent and swarm tool-result serialization to prefer compact evidence summaries before raw JSON

fixAdded regression coverage for packet schema validation, benchmark evidence-contract metrics, agent ingest summaries, and prompt serialization

v0.4.44

2026-05-050.4.44 Release

fixMade geotechnical document synthesis read an ordered whole-report outline before high-signal extraction rows so takeaways, risks, recommendations, and borehole evidence stay report-level first

fixPrevented figure, table, and borehole-log titles from replacing the report title in generated Geotechnical Intelligence Reports, including cached result rendering

featRebuilt borehole and ground-model report visuals with compact lithology-colored profiles, source-page legend rows, inferred-contact styling, and overflow-safe SVG layout

fixCurated noisy OCR material fragments out of the main material table while keeping page-level traceability in source evidence and Processing Audit

fixVerified the cached GeotechnicalInvestigationReport PDF output with Playwright screenshot and layout assertions for horizontal overflow and SVG text overflow

v0.4.43

2026-05-040.4.43 Release

fixIncreased the slow raster-page ingest recovery regression timeout so GitHub Actions can finish the same PDF/OCR-style path that passes locally on Windows

fixKept the benchmark, GroundModel readiness, provider-neutral evidence, and report filename behavior from 0.4.42 unchanged

v0.4.42

2026-05-040.4.42 Release

featAdded geotechnical report benchmark output for page cache reuse, estimated hosted calls, source-page traceability, retained signal counts, and GroundModel readiness gates

featAdded npm run benchmark:geotech-report as a local two-pass benchmark harness for the canonical GeotechnicalInvestigationReport PDF

fixStamped retained geotechnical materials, classifications, and parameters with provider-neutral source pages for hosted GLM and future BYOK providers

featAdded GroundModel calculation-readiness routing for bearing capacity, settlement, pile capacity, liquefaction, and slope stability

fixUpdated geotech analyze terminal and HTML report output to show deterministic workflow readiness before agents or users route evidence into calculations

fixRenamed default generated ingest and workspace HTML filenames from legacy review wording to report wording while preserving internal compatibility APIs

v0.4.41

2026-05-040.4.41 Release

fixRenamed the generated HTML ingest output to a Geotechnical Intelligence Report presentation, including CLI summary labels and browser report prompts

featReworked the report shell into a compact dark engineering dashboard with top navigation, cleaner hero hierarchy, dark evidence cards, and contained table/profile sections

fixTightened desktop and mobile layout constraints so status badges, source metadata, ground-model SVGs, borehole profiles, and engineering tables avoid overlap and horizontal clipping

fixKept Processing Audit and model-stage details below the engineering review surface while keeping the main report focused on decisions, evidence, confidence, and human verification

fixVerified the GeotechnicalInvestigationReport PDF report regenerated with all 34 pages processed and clean desktop/mobile screenshots

v0.4.40

2026-05-030.4.40 Release

fixStabilized PDF page evidence cache keys for generated per-page PDF payloads so repeated full-report ingest reuses native PDF page evidence instead of re-extracting regenerated pages

fixAdded regression coverage for synchronous and persisted async PDF page cache reuse when page payload bytes change between jobs

featUpgraded HTML ingest reports with status badges, searchable review filters, a schematic ground-model cross-section, source evidence actions, and a scroll-aware human review workflow bar

fixVerified the GeotechnicalInvestigationReport PDF rerun completed with all 34 pages as cache hits and no new cache stores while keeping the review-focused report presentation intact

v0.4.39

2026-05-030.4.39 Release

fixDisabled automatic page evidence cache reuse under Vitest unless a test opts in explicitly, preventing synthetic page fixtures from reusing stale cached extraction output across tests

fixKept production and local CLI cache defaults unchanged, so repeated real PDF ingest still reuses unchanged page evidence

fixVerified the full core test workspace after the cache-isolation fix

v0.4.38

2026-05-030.4.38 Release

featAdded a durable page evidence cache keyed by source file hash, page hash, page number, model version, preprocessing version, and schema version

fixReused cached geotechnical page evidence on reruns so repeated PDF ingest can skip duplicate OCR, GLM-OCR, GLM-5V, and page extraction work when inputs are unchanged

fixPersisted cache audit metadata through async ingest checkpoints, segmented child-job merges, final page audits, CLI summaries, and HTML report processing audit views

fixKept cache I/O best-effort so inaccessible local cache storage never turns a valid page extraction into a failed page

fixAdded regression coverage for cache key stability, corrupt cache misses, invalidation, sync ingest reuse, async job reuse, and report cache presentation

v0.4.37

2026-05-030.4.37 Release

fixRaised installed-CLI hosted beta defaults to 2,000 text, 600 vision, 600 layout, and 200 agent requests per day for image-heavy PDF development runs

fixRaised anonymous hosted beta defaults moderately while keeping stricter abuse protection than installed CLI traffic

fixHardened Geotechnical Intelligence Report borehole-profile inference to use retained inspection text and content chunks when structured parameter rows omit depth data

fixAdded regression coverage for profile generation from real report-style borehole schedule and conclusion text

v0.4.36

2026-05-030.4.36 Release

featRedesigned HTML ingest reports as a premium Geotechnical Intelligence Report with a sticky sidebar, executive facts, review actions, engineering insight cards, and source-evidence navigation

featAdded an evidence-first trust table for extracted and missing engineering parameters with source page, confidence, review posture, and evidence snippets

featAdded a lightweight borehole stratigraphy SVG view for retained borehole and layer evidence, with dashed boundaries for uncertain or missing intervals

fixMoved model-stage and raw page audit details into a collapsed Processing Audit section so the main report stays engineering-decision-first

fixAdded regression coverage for the premium report layout, trust layer, source evidence, processing audit, and borehole visualization

v0.4.35

2026-05-030.4.35 Release

fixKept GLM thinking enabled for final report synthesis as the first attempt, then retried once without thinking when Z.ai returns an empty/no-content assistant response

fixPrevented successful visual ingest runs from showing a GLM-5.1 synthesis failed warning when the fallback synthesis response is valid

fixAdded regression coverage for the hosted-beta synthesis fallback path so the report brief remains populated after an empty thinking-mode completion

v0.4.34

2026-05-030.4.34 Release

featAdded hosted GLM-OCR layout parsing with a separate layout quota bucket, configurable public limits, and developer key/IP bypass support

featWired geotechnical document ingest to prefer GLM-OCR layout text before GLM-5V visual extraction, then run GLM-5.1 synthesis for report takeaways, ground model, key parameters, interpretation, and limitations

fixRaised public hosted-beta CLI defaults to 300 text, 120 vision, 80 layout, and 40 agent requests per day while preserving tighter anonymous caps

fixRouted hosted-beta image-only, graphics-only, and text-unreadable report pages directly into structured GLM visual extraction when no accepted text exists, avoiding duplicate OCR-only vision calls before interpretation

fixMarked direct visual page extraction as vision-visual in page audits and forced manual review before approval so high-confidence image-only results do not silently auto-proceed

featImproved the HTML ingest report with report takeaways, grouped parameter tables, stage badges, an extraction overview strip, confidence meters, cleaner engineering brief text, and shortened table/page-card fragments for human review

featApplied a shadcn-style static report presentation pass so key engineering tables come before operational audit details, with raw page audit tables collapsed by default

fixPreserved engineering units such as kN/m3, t/m2, kg/cm2, and m/s in report presentation while still cleaning OCR-like spacing

fixAdded regression coverage for synchronous ingest, persisted async jobs, direct visual extraction, and report HTML rendering

v0.4.33

2026-05-010.4.33 Release

fixForced hosted-beta PDF page inputs through provider-safe raster images so geotech vision log no longer sends native PDF parts to the GLM vision model

fixApplied the same hosted-beta PDF raster path to CLI ingest and agent ingest tools while preserving native PDF pages for providers that can support them

fixRaised the public CLI hosted-beta vision burst window to four requests per minute so one-page image-only PDF workflows can complete OCR and structured extraction without self-rate-limiting

fixAdded regression coverage for forced raster PDF page inputs on digital-text PDFs

v0.4.32

2026-05-010.4.32 Release

fixDisabled GLM thinking mode for the hosted glm-5v-turbo vision path so borehole-log ingest and vision commands return final assistant content instead of empty completions

fixFixed the CLI vision retry backoff so recoverable hosted vision failures no longer let Node exit with an unsettled top-level await warning before the fallback attempt runs

fixAdded regression coverage for forwarding disabled thinking mode on GLM vision proxy requests

v0.4.31

2026-05-010.4.31 Release

fixDisabled GLM thinking mode by default in the hosted-beta proxy so low-latency text calls return final assistant content instead of spending tiny smoke-test budgets on reasoning tokens

fixAdded GEOTECHCLI_HOSTED_BETA_THINKING_MODE=disabled to the environment contract, with an opt-in path for future deeper reasoning tests

fixKept the Z.ai GLM model defaults and live release smoke checks aligned with the hosted-beta no-user-key contract

v0.4.30

2026-05-010.4.30 Release

featSwapped the strong-beta hosted AI defaults to Z.ai GLM, using glm-5.1 for text and agent reasoning and glm-5v-turbo for vision

fixKept the public CLI contract on hosted-beta so end users still get default AI access without bringing their own provider key

fixRetargeted the beta proxy, environment docs, release guardrails, and website copy around the server-side ZHIPU_API_KEY secret

fixDisabled the legacy Modal deploy workflow by default while preserving it as a manual historical fallback

fixAdded guardrails for GLM proxy model defaults and provider-neutral timeout and fallback wording

v0.4.29

2026-04-240.4.29 Release

featAdded the canonical GroundModel v1 contract for workspace analysis, binding boreholes, SPT tests, groundwater, lab parameters, monitoring series, and rejected observations into one engineering model

featAdded evidence-bound extraction for sampled CSV/XLSX data so analyze JSON and reports show source file, sheet, row, column, confidence, raw value, and normalized value

featAdded a deterministic GroundModel verifier for rejected SPT values, missing groundwater, missing coordinates, undeclared local CRS, unknown standard profiles, and duplicate SPT depths

featUpgraded geotech analyze --format html with GroundModel, verifier findings, and evidence table sections while keeping terminal output compact

fixUpgraded geotech agent --workspace so hosted-beta tasks receive a GroundModel/verifier summary without adding any Modal GPU calls to local analysis

v0.4.28

2026-04-240.4.28 Release

featAdded geotech analyze [workspace] for deterministic local project manifests with compact terminal output, JSON, and self-contained HTML reports

featAdded core workspace intelligence for file discovery, geotechnical classification, branch detection, recommendations, and ProjectManifest generation without hosted-beta GPU spend

featAdded lightweight CSV/XLSX schema inference for depth, coordinates, borehole/sample IDs, SPT/CPT, lab, monitoring, and signal-style columns

featAdded geotech agent ... --workspace <dir> so agent tasks can receive a compact local manifest summary instead of guessing directly from raw files

docsDocumented geotech analyze ., --branch, --standard, --json, and --format html while clearly marking deeper planner/calculation behavior as future roadmap layers

v0.4.27

2026-04-240.4.27 Release

fixReduced hosted Qwen Modal L4 admission pressure by defaulting concurrent inputs to 2, adding an explicit max-num-seqs guard, and keeping max containers at 1 for credit-safe operation

fixLowered hosted-beta proxy retry pressure with single-attempt vision and agent calls, fewer text retries, and separate per-minute limits for heavier vision and agent requests

fixSerialized more long mixed PDF ingest jobs on hosted beta so image-only report appendix and tail pages do not flood one GPU queue

fixAdded cheaper PDF retry behavior with deterministic partial extraction for text timeouts, skipped duplicate vision OCR on retry, and manual-review downgrades for repeated slow visual tail pages

featPrinted the exact browser-report command after plain persisted ingest results so users can immediately open the HTML review without rerunning the PDF

v0.4.26

2026-04-240.4.26 Release

fixFixed bundled strong-beta skills in global npm installs by trusting only first-party bundled-skill archives while keeping arbitrary outside ZIP imports blocked

featChanged long PDF ingest to live-wait by default with visible page progress, elapsed time, heartbeat/status, and failure counts, plus --background for detached jobs

featMade ingest HTML format save and open a browser report by default for run, wait, and result flows while --no-open keeps save-only automation clean

fixHardened PDF ingest with page-count fallbacks, zero-page job rejection, retryable 524/upstream-timeout checkpoints, image-heavy concurrency control, and completed-partial resume retry behavior

fixImproved report extraction by preserving user-declared report intent, limiting borehole appendix dominance, and rejecting impossible SPT values from standards references while retaining warnings as evidence

fixCleaned strong-beta drift across agent docs, handoff notes, changelog encoding, swarm final synthesis fallback, and effective config display for provider defaults

v0.4.25

2026-04-230.4.25 Release

featAdded hosted-beta long-PDF segmentation for geotechnical report ingest, using a 60 effective-page best-result window with linked sequential packets and one merged final result

featAdded --page-range <start:end> to geotech ingest so targeted report subsets can be reviewed, rerun, and exported without processing the whole PDF

featExtended the HTML ingest report to show segment execution, selected page-range context, and merged packet outcomes inside one premium review surface

fixKept the shipped hosted-beta path on the non-quantized default Qwen model while forcing segmented long-report execution onto extraction concurrency 1 for better queue stability on L4

fixHardened parent and child persisted ingest jobs so segmented report packets checkpoint independently and merge back into the parent result with original page numbering preserved

v0.4.24

2026-04-230.4.24 Release

fixFixed the Modal deploy workflow so blank repository variables now fall back to the intended defaults instead of crashing the hosted-beta launcher during GitHub Actions deploys

fixReduced hosted-beta L4 startup memory pressure by defaulting the vLLM compilation profile to cudagraph_mode NONE, enabling PyTorch expandable segments, and stripping stale num-gpu-blocks override startup args that can force KV-cache OOM

fixWired the new Modal compilation override through the deploy workflow and runtime notes so the hosted Qwen path can be tuned safely without turning on a baseline warm-container cost increase

fixUpdated the strong-beta release surfaces for 0.4.24 so package metadata, lockfile pins, changelog notes, and the website release contract stay aligned with the shipped hosted-beta hotfixes

v0.4.22

2026-04-210.4.22 Release

fixUnified the live single-agent and swarm tool bootstrap so both runtimes register the same filesystem, shell, data, deliverable, and skill tools instead of drifting when a tool depends on side-effect imports

fixAligned swarm with the same deterministic intake screen and first-turn hosted-beta fallback used by the main agent, so under-specified requests and provider warmup/outage cases now resolve more consistently across both execution paths

fixUpdated the strong-beta release surfaces together for this versioned reliability pass, including package metadata, README, website docs, website changelog, homepage bundled-skill copy, and release-consistency checks

featAdded a self-contained HTML ingest report for geotech ingest so borehole packets and broader geotechnical reports can be reviewed as a polished engineering brief instead of raw JSON alone

featAdded public strong-beta docs and README examples for report ingest, resumable long-PDF jobs, and HTML report export using real geotechnical report workflows

fixContinued hardening the long-PDF ingest path around normalized page routing, async job orchestration, and broader geotechnical report extraction so mixed engineering PDFs behave like a first-class workflow

v0.4.21

2026-04-200.4.21 Release

securityAdded a dedicated proprietary-internals refusal policy across the main agent, swarm, and exported legacy multi-agent orchestrator paths so hidden prompts, internal instructions, repo layout, and protected source-file requests are blocked before any model call is made

securityHardened the agent filesystem and shell sandbox against geotechCLI repo-internal disclosure by blocking protected source/build/test areas and rejecting broad repo-root enumeration when the CLI is run from its own source checkout

fixKept normal user flexibility intact by preserving high-level product explanations, ordinary project subdirectory scanning, and the existing opt-in command runner behavior while adding regression coverage for the new protection boundaries

v0.4.20

2026-04-200.4.20 Release

fixAdded a repo-level contributor guide so maintainers and coding agents can follow the current strong-beta architecture, release workflow, and verification expectations more consistently

fixRemoved a stale internal swarm prompt reference to the deleted local integration surface so the shipped multi-agent guidance matches the real toolset and beta scope

fixKept the strong-beta build green after the internal agent-guidance cleanup so the CLI, core package, and web app stay aligned on the new release

v0.4.19

2026-04-200.4.19 Release

fixTrimmed an unfinished local integration surface from the shipped beta so the public CLI only exposes features that are currently supported end to end

fixRemoved the related internal tool wiring and exports so the packaged runtime and public API surface stay aligned with the current beta scope

fixCleaned docs, pricing copy, README content, and release notes so the website reflects the supported feature set more accurately

v0.4.18

2026-04-200.4.18 Release

featAdded a dedicated geotech skill section to the public docs with concrete list, show, validate, run, and agent-session examples so bundled skills are documented as a first-class strong-beta surface

featAdded a homepage bundled-skills showcase on the beta site, including a direct docs link and terminal examples for direct CLI use and the per-session --skills agent opt-in path

fixUpdated the public beta capability messaging to show the bundled strong-beta skill catalog more explicitly instead of leaving skills buried inside release notes or command help

v0.4.17

2026-04-200.4.17 Release

fixReplaced bundled skill ZIP extraction with a pure JavaScript path inside @geotechcli/core so first-use skill bootstrap no longer depends on external tar or python executables just to list or inspect installed skills

fixPreserved the strong-beta skill catalog and the new --skills agent session opt-in while removing the Windows-specific archive extraction footgun that could block fresh local skill discovery

fixKept archive path validation in place so bundled skill imports still reject unsafe ZIP paths during bootstrap and nested archive discovery

v0.4.12

2026-04-190.4.12 Release

fixAdded strong-beta release guardrails so public beta pushes now require a new shared version and verify GitHub, npm, Cloudflare, and Modal release surfaces together

featRemoved the Modal deploy path filter so versioned strong-beta releases always refresh the hosted Qwen runtime instead of silently skipping the Modal workflow

featExtended the strong-beta skill runner with compatibility adapters for the remaining EPB tunnel workspace bundles so tunnel skills now share the same trusted install, sandbox, and artifact pipeline as the rest of the catalog

featApproved the remaining EPB tunnel skills for strong-beta execution after end-to-end certification against their bundled example inputs

fixFixed skill run workspace naming so back-to-back skill executions cannot collide in the same millisecond and corrupt output bundles

fixExpanded skill runtime regression coverage so the full tunnel workspace compatibility set stays aligned with the strong-beta approval catalog

v0.4.11

2026-04-170.4.11 Release

fixRaised hosted-beta agent timeout budgets to better match real Modal cold starts so the CLI is less likely to fall back before the Qwen server is actually ready

fixChanged first-turn fallback messaging to explicitly say when the Modal.com GPU is warming up or the timeout budget was exceeded instead of implying the hosted stack is simply unavailable

featAdded live waiting hints for hosted-beta sessions so geotech agent and geotech chat can show a clearer Modal.com GPU warmup message while users wait

fixLimited hosted agent proxy calls to a single upstream attempt instead of retrying cold starts, reducing wasted L4 GPU time and keeping timeout behavior more honest for the current budget

fixAdded regression coverage for Modal warmup and timeout fallback wording plus the no-retry hosted agent proxy path

v0.4.10

2026-04-170.4.10 Release

featAdded a terminal-rich text renderer for geotech agent, swarm, and chat answers so sections, bullets, tables, and inline tool names render as a cleaner engineering brief instead of raw markdown

featRefined the interactive plot viewer with a lighter product-style surface, scrollable legends, contained labels, and inside-only zoom so the old thick bottom slider no longer crowds the plot area

fixShortened chart subtitles and adjusted chart spacing so the actual engineering plot keeps priority over surrounding viewer chrome on narrower screens

fixSkipped mixed-unit overview charts for generic CSV and workbook plots unless the overlaid series share a clear unit signature

fixPromoted CPT templates to true inverted depth profiles, split pile browser plots into separate force and stress views, and corrected liquefaction chart domains so factor-of-safety charts are no longer stretched by blow-count ranges

v0.4.9

2026-04-170.4.9 Release

fixGeneric hosted-beta transport failures like Hosted beta AI request failed, empty upstream content, and fetch failed are now treated as temporary first-turn availability issues instead of surfacing a raw chat error and then continuing down a second answer path

featAdded a live terminal status controller for geotech agent and geotech chat so users see Terzaghi actively thinking, calling tools, and reviewing results while the response is being prepared

fixKept the underlying agent workflow unchanged for evidence-backed requests while making the waiting experience feel active and more trustworthy during long hosted responses

fixBroadened tunnelling intake detection so UCS, water inflow, face, and machine-selection prompts now short-circuit into the data-requirements screen even when users do not explicitly say TBM or tunnel

v0.4.8

2026-04-170.4.8 Release

featReplaced the narrow foundation-only shortcut with a catalog-driven geotechnical intake layer that screens multiple analysis families before calling the hosted model

fixUnder-specified foundation, soil classification, and liquefaction prompts now return immediate engineering data requirements instead of burning a long hosted-beta round trip

fixProject metadata and notes alone no longer suppress agent intake; only real evidence such as soil profiles, datasets, derived parameters, or active analysis context can bypass it

fixAdded regression coverage to keep metadata-only project sessions from slipping past intake while still allowing evidence-backed project context to reach hosted beta normally

v0.4.7

2026-04-170.4.7 Release

fixAdded an immediate deterministic screening answer for under-specified foundation and soil-profile agent prompts so the CLI responds instantly instead of waiting on hosted-beta just to request missing data

fixImproved the hosted-beta fallback for the same foundation-screening prompts so transient provider issues now return a useful engineering checklist instead of the generic no-fallback message

fixSwitched hosted beta to the native vLLM OpenAI-compatible server on Modal so text and image chat completions share the same contract and the broken modal-http invalid function call path is removed

fixKept hosted beta on Qwen/Qwen3.5-9B as the shared hybrid multimodal model instead of drifting to a separate Qwen2.5-VL default

fixCorrected the Modal vLLM launcher to pass --limit-mm-per-prompt as JSON so the current vllm serve CLI accepts multimodal limits during deploy

featExpanded the Modal deploy workflow to watch hosted-beta contract files and added a post-deploy /health smoke check to catch drift immediately

featReplaced ASCII-first plotting with an interactive browser plot viewer for geotech viz and command-level --plot flows, while keeping terminal fallback available

vstrong-beta-qwen-migration

2026-04-16Qwen on Modal

featMigrated hosted beta backend from GLM (Zhipu/Z.AI) to Qwen/Qwen3.5-9B served on Modal.com with NVIDIA L4 GPU

featModal deployment auto-scales to zero after 10 minutes idle to conserve GPU credits

featHosted beta text and vision defaults unified to Qwen/Qwen3.5-9B

fixUpdated proxy, CLI defaults, docs, privacy, and website copy for the new model and provider

vstrong-beta-model-refresh

2026-04-14Hosted Model Update

featHosted beta text default switched to glm-4.7-flash to reduce shared credit burn during strong beta

featHosted beta vision default switched to glm-4.6v-flash for lower-cost image analysis during strong beta

fixHosted proxy allowlist, CLI defaults, and docs were aligned to the new flash model IDs

vstrong-beta-wave2

2026-04-03Hosted Qwen Beta

featHosted beta gateway enabled for strong-beta with Qwen/Qwen3.5-9B on Modal

securityProxy now validates requests, enforces model allowlists, and applies server-side rate limits before calling upstream

featCLI default provider switched to hosted-beta so users can try AI commands without bringing their own key

featWebsite, docs, and privacy copy updated to reflect hosted beta access with no-signup limits

vstrong-beta

2026-04-03Strong Beta Branch

featPublic beta branch introduced for safe Cloudflare deployment trials

featWebsite messaging rewritten around strong beta: deterministic CLI live, hosted anonymous AI coming in a later wave

fixSignup, checkout, usage, webhook, and hosted proxy endpoints disabled until the beta gateway is ready

fixCLI AI flows now use the user configured provider directly in Wave 1, without fake registration walls

featHosted beta defaults are Qwen/Qwen3.5-9B for both text and vision

v0.4.4

2026-04-140.4.4 Release

fixDeepened hosted-beta upstream retry and backoff behavior again so transient free-model overload windows are less likely to surface as immediate CLI failures

featAdded a local heuristic fallback for geotech ai-classify so descriptive soil classification can still return a USCS-style result when hosted-beta text is temporarily unavailable

fixAdded deterministic fallback behavior for the first agent turn so provider saturation now yields a useful engineering limitation analysis instead of a blank agent failure

fixAdded focused regression tests for the new ai-classify and agent fallback paths before releasing the patch

featRaised hosted-beta limits for installed geotechCLI clients while keeping stricter anonymous caps in place for the public beta proxy

fixAdded retry and backoff handling for transient upstream 429 and gateway overload responses so simple AI commands recover more gracefully

fixExtended hosted-beta timeout budgets for text, vision, and agent requests and translated raw aborts into clearer timeout messages in the CLI

fixSeparated daily usage fingerprints by client mode so geotechCLI traffic no longer burns through the same low anonymous bucket

fixAdded focused hosted-beta regression tests covering client-mode limits, upstream retry behavior, and timeout handling before shipping the patch release

featExpanded terminal plotting so classify uscs, liquefaction, and pile now support direct engineering ASCII charts through the shared --plot flag

featAdded geotech viz engineering presets for Mohr circle and Atterberg plasticity charts with field-friendly CLI inputs

featAdded geotech viz file templates for compaction curves, grain-size distribution curves, and CPT-style plotting to speed up daily engineering review

featCommitted new visualization showcase files for compaction, gradation, and CPT workflows and kept the docs in sync with the CLI surface

securityClosed agent sandbox escape paths, gated command execution more tightly, and hardened hosted-beta request handling for production-facing strong-beta use

fixCorrected slope seismic regression behavior and locked the deterministic engineering suite back to a passing state

featAdded additive case-file persistence, deterministic report assembly, conservative evidence records, and agent deliverable tools for report and export generation

featEnabled PDF and DOCX exports for deterministic stored-case reports using dedicated layout generators

featAdded terminal visualization with geotech viz plus committed CSV and Excel showcase samples for demos and quick plotting

featBorehole log vision can now process multi-page PDFs page by page and merge the extracted log output into one result

fixAdded settlement trough and forgiving through aliases so the Peck plotting command matches the public examples and common user phrasing

fixExport examples and committed mock datasets were added so GeoJSON, DXF, and CSV flows can be smoke-tested reliably

securityUpdated Next.js to 15.5.15 and the Vitest toolchain to 4.1.4 so the release workspace installs cleanly with no npm audit vulnerabilities

fixAligned GitHub Actions and Cloudflare release jobs to Node 22 and added the direct esbuild install required by the OpenNext Cloudflare bundler

fixAligned the public release surface on Qwen3.5-9B defaults across docs, CLI, and website copy

v0.2.0

2026-03-30Stability + Safety

securityFilesystem sandbox and shell command hardening for agent tools

featPile capacity, slope stability, and lateral earth pressure modules added to the deterministic core

fixBearing capacity water-table correction implemented for shallow foundations

featPersistent CLI usage tracking and email verification flow added

feat--quiet and --dry-run wired into the CLI for safer scripted workflows

fixExpanded guardrails, standards coverage, and tool registry breadth for agent execution

v0.1.0

2026-03-26Initial Release

featBearing capacity calculation (Terzaghi, Meyerhof, Hansen, Vesic)

featLiquefaction triggering analysis (Boulanger & Idriss 2014, NCEER)

featRock & soil classification (RMR89, USCS, Q-system)

featTBM performance prediction, type selection, cutter wear estimation

featAI vision analysis: core box, hybrid RMR, sensor data, borehole logs

featMulti-agent orchestrator (Geo, Tunnel, Hydro, Seismic, Slope, Foundation agents)

featAI-powered report generation from analysis data

featNatural language soil classification

featGBR document Q&A with vision model

featExport to GeoJSON, DXF (AutoCAD), and CSV

featLLM-agnostic provider layer: Qwen, OpenAI, Anthropic, Hugging Face, and self-hosted backends

featAnti-abuse metering: 5 free AI calls for unregistered users, IP fingerprinting

feat--json, --verbose, --plot, --output global flags on every command

securityAPI keys never logged, echoed, or included in error messages

securityJSON output auto-redacts all sensitive fields