- Corrected the live public PyPI boundary from v1.11.1 to v1.18.0 after post-release verification, and moved exact GitHub install links to the immutable v1.19.1 patch artifacts. No Evidence, authority, Review Graph, or Behavioural Risk semantics changed.
-
Reframed AET as a local Evidence Plane for coding Agents and rebuilt the bilingual README, static/dynamic architecture, exact install paths, case library, and 30-second introductions around the complete current product.
-
Added Review Graph: a Git-bound Python Code Graph composed with Bundle evidence and Improvement controls. The default Agent input is a 6,505-byte, hash-bound root slice with bounded one-hop expansion; stale or tampered packages return an
UNKNOWNstop condition. In one frozen case this reduced the minimum raw review material by 23.2%; it is not a universal token or model-quality claim. -
Added
aet review-graph build|validate|open|expand|export-compat, four strict Schemas, a non-overwriting Review Package, human Mermaid projections, two read-only MCP tools, fail-closed scope and snapshot checks, and explicit compatibility export for legacy Agent context files. -
Added the AET Lab
aet risk diagnosesurface with strict Risk Policy and Diagnosis v1 schemas, deterministic three-factor findings, same-context pathways, evidence citations, four-state semantics, andPROPOSED-only interventions. It performs no model training, model calls, or automated action and produces no single aggregate risk rating or internal-motive claim. -
Added the gated experimental
aet risk forecastsurface. Unsupported, out-of-domain, leaking, or insufficiently calibrated pathways remainUNKNOWN. Forecast promotion is hard-disabled as research-only in this release, so self-declared calibration metrics cannot enable a prediction. -
Added a read-only MCP diagnosis tool, an optional Evidence Atlas behavioural risk projection, Python/TypeScript protocol contracts and local validators, offline evaluation fixtures, and Codex/Claude Code parity coverage. Evidence Atlas now exposes eleven fixed Perspectives; an absent risk diagnosis is an explicit
UNKNOWNview. -
Added a diagnosis-only public release gate derived from nine programmatically scored AgentDojo runs at a pinned MIT-licensed upstream commit. Minimal action/outcome summaries retain source and event hashes without redistributing prompts or benchmark solutions; human validation and forecast eligibility are explicitly not claimed.
-
Corrected distribution documentation: the exact current package is published through GitHub Releases, while public PyPI remains on the older v1.18.0 feature set.
- Added an installed-wheel
aet demo stale-proofpath that runs a real standard-library test, records Proof with the existing Quick API, verifiesEXACT_MATCH, mutates a declared source, and verifiesRELEVANT_FILES_CHANGED. - Reframed the first-run experience around proof-carrying workflows for coding agents, with focused English and Simplified Chinese documentation, a static no-tracking site, community guidance, and reproducible visual assets.
- Added strict Skills catalog validation, a read-only GitHub Action template, launch-readiness checks, and privacy-preserving growth snapshots. All external publication, repository settings, and outreach remain manual.
- This activation and distribution release does not change core Evidence authority semantics: tests and Proof remain source-bound, missing facts remain explicit, and human adoption and release authority are unchanged.
- Added Evidence-Guided Planner as a read-only Host boundary:
deterministic Planning Request and Context construction from validated
Bundle, optional matching Atlas, current hash-bound Source, existing scope,
and explicit budgets; Host reasoning returns strict Candidate JSON, and AET
alone validates the portable
PROPOSEDPlan. - Added six strict Planning v1 Schemas, fail-closed identity/reference/path and
protected-scope validation, explicit
READY_FOR_HUMAN_REVIEW,NEEDS_EVIDENCE,PARTIAL,BLOCKED, andSUPERSEDEDstatus, four Edit dispositions, counter-evidence/conflict/UNKNOWN preservation, and guardedBOUNDED_COMPLETEcoverage. - Added
aet plan context|validate-candidate|validate|show|explain|trace|gaps| export-skill|verification-handoff, deterministic Markdown and portable Plan packages, single-Plan Host Skill export, and external-diff handoff that leaves ProofUNKNOWNandPENDINGinstead of executing it. - Added the thin
/aet-planHost Skill, eight bounded Planning MCP tools, Python SDK helpers, andloadPlan,queryPlanEdits, andvalidatePlanReferencesin the existing TypeScript Bundle SDK. No model client, Agent loop, source editor, or Executor was added. - Added three reproducible end-to-end examples and a separately frozen, 20-case, three-group localization benchmark. Release language is deliberately “bounded localization”; it does not claim to find every modification point.
- Added English and Simplified Chinese product navigation, complete Planner, Helper, Schema, architecture, security, SDK, MCP, example, and verification handoff documentation.
- Added a reproducible real Codex
gpt-5.6-solAET self-review with separate source-only, v1.16 evidence-only, and v1.17 validated-Plan observations. The checked-in Gold, raw structured outputs, deterministic scorer, and result report eight independent scope-localization metrics without a holistic score. In this one-run-per-group case, v1.17 raised production-path decision precision from 44.44% to 100%, linkage coverage from 33.33% to 100%, andUNKNOWNpreservation from 0% to 100%; path/test recall was already 100% in source-only and did not improve. - Rebuilt the bilingual README Planner story, dynamic/static architecture, real-case screenshots, project panoramas, and exact 30-second silent H.264 introductions. SHA-256 manifests bind the generated workflow, review contact sheets, screenshots, panoramas, and videos.
- Added the deterministic Evidence-Grounded Improvement Layer: validated
Portable Evidence Bundle records become bounded Improvement Issues,
Constraints, human reports,
PROPOSEDAgent tasks, Verification Contracts, and outcomes without editing code or granting recommendations evidence authority. - Added
aet improvement doctorandaet improve <bundle>|prompt|validate|verify|compare, strict grounding/scope/strength and anti-gaming checks, protected-path enforcement, Proof/Freshness binding, and before/after comparison that retainsPASS,FAIL,UNKNOWN, andNOT_APPLICABLE. - Extended Evidence Atlas from eight to ten fixed Perspectives with
Improvement Chain and Regression Lineage. Bundle v1 has no independent
Improvement records, so these views remain explicitly
UNKNOWNuntil source-bound records exist; prompts never write back as Evidence. - Added a reproducible project-review case in which an empty tool result is
incorrectly emitted as “no issues.” A real failing regression becomes
counter-evidence for
IMP-001, grounding the human report, one-file Agent scope, verification command, and Claim/Improvement Chain diagrams. - Fixed Mermaid rendering so unsupported Claims use a distinct red dashed class instead of the supported-Claim style.
- Rebuilt the English and Simplified Chinese README product narrative, commands, architecture and process diagrams, real-case GIF/PNG/SVG assets, Evidence Atlas visuals, and silent 30-second H.264 introduction videos with refreshed SHA-256 manifests.
- Added Evidence Atlas, a deterministic canonical Evidence Graph derived from
Portable Evidence Bundle v1 records. Authoritative nodes and edges retain
field-level source references; stale Proofs cannot validate current Claims,
counter-evidence remains visible, and missing facts remain
UNKNOWN. - Added eight fixed Perspectives for Claim Chain, Investigation Flow, Change Scope, Verification Coverage, Evidence Data Flow, Integrations, Conflicts, and Freshness. Every node receives deterministic complexity evaluation; typed recursive decomposition, canonical-node deduplication, cycle stops, and hard depth/node/child/diagram budgets keep diagrams bounded.
- Added safe Diagram IR and Mermaid projection, structured Markdown and JSON, strict validation, and a completely offline three-column Viewer with local Mermaid, search, filters, supporting/counter/stale path highlighting, evidence details, and recursive parent/child navigation.
- Added
aet atlas build|validate|view|export|query|explain|diff, optional Python and TypeScript graph APIs, and eight bounded read-only MCP graph tools. Incremental rebuilds reuse validated unchanged Perspectives and unaffected recursive subgraphs; Atlas comparison reports Claim, Freshness, conflict, and unknown changes without a holistic trust score. - Added JSON Schemas, security and failure fixtures, protocol tests, and a reproducible AET self-review Bundle whose real Claim Chain and offline Viewer drive the README example.
- Updated the English and Simplified Chinese product showcase with synchronized static Evidence Atlas architecture, a six-state real Viewer GIF, 30-second bilingual H.264 walkthroughs, a SHA-256 media manifest, and a GitHub-rendered Mermaid self-review example.
- Added host-neutral Codex and Claude Code Run Normalizers with stable record identity, content hashes, tool-call/result linking, incremental ingestion, generation boundaries, and explicit diagnostics.
- Added the strict Observation → Evidence Candidate → Verified Evidence boundary. Agent statements, reasoning, and recorded tool output cannot become reproduced proof without deterministic workspace-bound verification.
- Added the read-only Portable Investigator with primary and competing
hypotheses, disconfirming search, privacy policy, budgets, stop conditions,
immutable ledger output, and fail-closed
UNKNOWNhandling. Explicitly authorized AET Proof receipts can now verify matching command Candidates against the declared read-only workspace; Freshness drift preserves the historical result while preventing promotion to current proof. - Added Portable Evidence Bundle v1: self-describing Index/Core/Archive JSON and JSONL, deterministic Markdown, content-addressed Blobs, Freshness, counter-evidence, conflicts, diagnostics, canonical hashing, strict schema validation, and optional structured Review Result validation. Reviewers do not need AET or an SDK installed.
- Added bounded CLI and MCP operations plus optional Python and TypeScript Bundle SDKs. The TypeScript package ships a Node 20-compatible JavaScript distribution, deep-frozen trusted handles, strict duplicate-key parsing, and adversarial tests for forged claims, evidence, enums, links, and Blobs.
- Added ten deterministic consumption scenarios and independently rescorable
prompt-only measurements for Codex CLI 0.144.1 /
gpt-5.6-sol, Hermes Agent 0.17.0 /kimi-k2.6, and Ollama 0.32.3 /qwen3:8b. Every consumer produced strict JSON covering all ten scenarios with 62PASS, 38NOT_APPLICABLE, zeroFAIL, and zeroUNKNOWN; these are bounded synthetic-fixture interoperability results, not general accuracy or trust scores. - Rebuilt the English and Simplified Chinese README showcase around the current portable evidence architecture, including synchronized static SVG/PNG panoramas, validated 5.75-second bilingual GIF workflows, exact 30-second bilingual H.264/AAC product videos, and a SHA-256 media manifest.
- Added the bounded
OptimizationCandidateentry contract. It requires multiple independent tasks or an explicit current high-severity deterministic failure, always requires isolated evaluation, and cannot modify a Skill, adopt a candidate, or publish a release.
- Added the four bounded AET Quick surfaces: Check, Scope, Proof, and Fresh,
while preserving the legacy 1.x
audit,review,trace, receipt, and Lab commands. - Added host-neutral Intent, Investigation Ledger, Investigation Contract, Investigated Finding, and Command Budget contracts plus deterministic grounding, permission, source, search-scope, recorded-conflict, contract-shape, and stop-policy validation.
- Added relevant-file, artifact, runtime, and lockfile binding to minimal Quick proof receipts, with seven explicit freshness states.
- Added dedicated portable Skills and deterministic language routing: a Chinese slash-command request receives natural Simplified Chinese narrative; every other request defaults to English. Machine statuses and evidence references remain language-neutral.
- Rebuilt the English and Simplified Chinese product entrypoints around AET Quick, with synchronized static/animated architecture, bilingual introduction videos, and the existing commit-locked Repository Audit Showcase retained as an explicit AET Lab case library.
- Added an opt-in AET Lab comparison across pure rules, one-shot LLM,
investigated AET, and a distinct Grounding-aware Agent configuration plus the
shipped Grounding Validator. The eight-case, two-repetition result publishes
64 privacy-reviewed normalized Runs for independent rescoring and records
recall, false discovery proportion, ungrounded conclusions, tool calls, wall
time, and Tokens while leaving unmeasured human-review and understanding
metrics
UNKNOWN. - Added a tracked 30-sample local performance report with raw Check, Scope, and Fresh timings, nearest-rank P95, environment, repository size, and explicit limits.
- Added the Repository Audit Showcase for commit-locked SWE-agent, Google ADK,
and OpenHands checkouts, with independent
repository-audit-profile/v1profiles and theaet audit <case> --repo <checkout>CLI. - Added deterministic, static-only evidence collection and findings. Upstream
code and tests are never executed; LLM narration is off and cannot change a
Finding;
PASS/FAIL/UNKNOWN/NOT_APPLICABLEremain authoritative. - Added profile, result, and evidence-manifest JSON Schemas, shared machine artifacts, and complete English/Simplified Chinese summaries, Markdown, HTML, Agent-flow SVGs, and evidence-chain SVGs.
- Locked each upstream commit and License boundary. OpenHands
enterprise/**content is prohibited, and its externally versioned Agent core remains an explicitUNKNOWNrather than a guessed local capability. - Added bounded-path, symbolic-link, stale-output, Schema, false-positive, runtime-budget, CLI-compatibility, and no-source-redistribution regression coverage.
- Reworked the bilingual entrypoint around one concrete failure: a passing test log becomes stale when the workspace changes. Added a runnable stale-proof demo and case study that preserve historical execution success while failing current freshness.
- Published the AET 1.x stability contract and moved three focused contribution paths near the top of both READMEs.
- Added a focused PyPI project page, complete package metadata and an OIDC Trusted Publishing workflow that promotes only the exact wheel already built and hash-verified by CI.
- Replaced the universal six-pair Real Host release constant with a
pre-registered
gate-plan/v2. Plans bind Claim, risk, Candidate, Runner, configuration, Scorer, Task and Fixture bytes; R3/R4 plans include an exact directional power analysis and risk-declared effect assumptions. - Split observed Suite semantics: Core is a zero-hard-regression, all-candidate-tasks-pass retention contract; Validation and Held-out use explicit one-sided exact paired objectives plus MCID. Legacy commands retain their historical two-sided McNemar behavior.
- Added dependency-free Bonferroni group-sequential alpha spending, efficacy, hard-regression, infrastructure, mathematically safe futility and maximum-N stopping reasons. Ordinary fixed-sample p-values cannot be repeatedly peeked.
- Added full observed-task preflight before the first Host call, append-safe
replay checkpoints, explicit exact
--resume, and same-run replay reuse for Tournament. Sleep no longer performs a duplicate Replay before Gate, and Held-out is not opened after a preceding objective fails. - Added strict
gate-history-registry/v1and planning-only history assessment. Entries require verified PASS provenance, exact identity, count consistency and deduplication. Drift, discount caps and leave-one-release-out sensitivity are reported; historical observations never enter PASS and never lower the fresh-pair planned maximum. - Reduced the portable root Skill from 262 lines / 14,401 bytes to 99 lines / 5,926 bytes by routing to on-demand delivery, provenance, quality, evolution and security references while preserving default-off activation and immutable authority boundaries.
- Removed repeated CI test subsets, added concurrency cancellation, and changed Release to verify and promote the exact commit-bound CI wheel instead of retesting and rebuilding a different artifact.
- Reworked both READMEs and bilingual architecture diagrams around conditional evidence budgets, fresh-only Gate decisions, planning-only history and the asymmetric human Adoption boundary.
-
Added explicit lossless Trace reuse with
aet trace --reuse-if-fresh. Reuse never executes or falls back to execution and requires an exact non-secret argv digest, safe rendered argv, proof binding, declared-artifact set and bytes, stdout/stderr log bytes, successful source status, and current full Git workspace snapshot. Any argv redaction disables reuse rather than persisting a guessable secret digest. Canonical Trace JSON is protected by an adjacent integrity seal; validator FAIL/UNKNOWN is propagated into the Trace summary, Run Gate, and reuse decision. Missing legacy fields, tampering, or drift fail closed. -
Added
aet evidence receipt, a compact hash-bound index for canonical Audit, Review, Trace, and Evidence Pack JSON. Canonical evidence remains unchanged; Agents can consume the receipt without loading full findings or embedded artifacts into context. Receipts independently recompute live workspace freshness and cannot overwrite their canonical source. -
Removed redundant in-request snapshot work from Trace and Run initialization, and made identical Run artifact attachments idempotent. Persistent snapshot caching remains intentionally absent because it cannot prove untracked-file freshness without rereading content.
-
Generalized release evidence directories and candidate identifiers from the source version, removing v1.9 workflow and manifest path hard-coding.
-
Classified releases explicitly as
deterministicorgovernance-adoption. Deterministic runtime/evidence releases record the Real Host Gate asNOT_APPLICABLE; only adoption releases that claim changed Agent behavior require the complete commit-bound paired Gate. The workflow rejects both a missing adoption Gate and an irrelevant Gate attached to a deterministic release, and publishes the disposition asrelease-evidence.json. Classification is a tracked, base-tag and Diff-digest-bound contract: behavior-sensitive paths require exact reviewed exceptions with deterministic proofs when a Gate is not applicable. Releases retain the contract, commit-bound verification, evidence disposition, and (for adoption) verified Gate manifest as durable assets. Adoption contracts bind structured claim IDs and covered Suite IDs to the exact Candidate SHA verified by that manifest. -
Made the portable Skill explicitly opt-in and default-off: installation no longer implies authorization for routine coding or review, activation is scoped to the current user-requested task, and real-host evaluation/evolution requires separate explicit intent. Added bilingual project-fit and cost guidance distinguishing removable orchestration overhead from the rollout and suite coverage required for adoption-grade statistical confidence.
-
Reworked the English and Chinese project entrypoints around AET's current position as an evidence-driven control plane for Agent-engineered repositories. The new narrative leads with the verified v1.9 real-host Gate, makes the Evidence → Quality → bounded Evolution → human-authority model explicit, sharpens toolchain differentiation and trust boundaries, and replaces the previous Mermaid renders with bilingual, editable dark-console architecture diagrams. The diagrams now separate Evidence Pack inputs from independent provenance stores, Quality regression staging from governance asset adoption, and audit-rule-only Shadow from the general adoption path.
- Added a deterministic Quality layer:
aet quality diagnosemaps structured failure phenomena to explicit owner, action, confidence, review route, and bounded repair surface without rewriting source status;quality promotestages confirmed badcases as canonical, validation-only Learn Task v2 candidates with immutable diagnosis provenance and deduplicated support. - Added Learn Task v2 contracts for fixture integrity, ordered tool calls,
argument constraints, proof/artifact requirements, command and change
budgets, and deterministic suite verification. Observed replay now reports
repeated-run any-success, all-success, Wilson intervals, paired McNemar
statistics, and explicit
INCONCLUSIVE/INFRASTRUCTURE_ERRORstates. - Hardened observed execution by reporting scripted network isolation as
PARTIAL, rejecting unsupportedenforced-denybefore execution, copying and hashing fixtures without following links or special files, and accepting Trace credit only through the injected./.aet-rollout/bin/aet tracepath. Snapshot state is independently recomputed without self-referencing the Trace JSON, and declared artifacts/logs are bound to real workspace files by source hash, size, fixed log path, independent redaction, and freshness. Outer child argv, Trace argv, and the intent proof command must match exactly; proof evidence must be an array. - Clarified the privacy boundary between private raw rollout material and
Evidence Only exports, including explicit environment-name allowlists for
real-host runners. Process adapters require both Task permission and the
inherit_homeswitch forHOME; the scripted adapter uses only the Task allowlist. Environment inheritance never makes secret values public evidence. - Added deterministic business-flow fixtures and separated core, validation, and held-out real-host proof suites. A manual Codex workflow can produce a commit-, version-, candidate-, task-, fixture-, and raw-gate-bound release artifact; the release workflow reconstructs and verifies it. The v1.9 tag was locally gated with authenticated Codex CLI 0.144.1: core, validation and held-out each produced 6/6 candidate successes, 0/6 baseline successes, zero infrastructure failures and exact paired p=0.03125.
- Pinned the release runner to
@openai/codex@0.144.1; process runners now capture and cache the canonical--versionoutput, bind runner name/version through raw manifests, observed replays, observed Gates, and release manifests, reject blank/unknown version probes, and reject mismatched release provenance. - Reframed AET as an evidence-driven Agent engineering quality and control layer, refreshed the bilingual user guide and architecture, and documented what AET deliberately does not provide: general benchmarking, LLM-Judge-led scoring, automatic semantic RCA/clustering, automatic repair/adoption, or an online ticket and business-metrics platform.
- Generalized Evidence-Gated Evolution from a Skill-only pipeline into six Constitution-bound targets: Skill, audit rule, audit profile, review policy, Trace validator, and triage policy. Legacy v1 Skill candidates are upgraded in memory; new candidates use a hash-bound Candidate IR v2.
- Added a non-executable declarative audit RulePack, rulepack identity in Audit reports, reproducible Audit Feedback, 30 core / 15 validation / 15 held-out / 10 adversarial audit tasks, baseline/candidate fixture replay, and monotonic multi-dimensional Gates.
- Added Shadow Audit that never changes official findings or exit status. Audit-rule adoption additionally requires 20 shadow runs across five repository fingerprints and three dates, every new finding confirmed, and zero confirmed false positives.
- Added bounded audit profiles, monotonic review policies, built-in JUnit/SARIF/coverage/JSON Trace validators bound to fresh declared artifacts, and triage policies that can reorder but never hide or rewrite findings.
- Fixed fenced Markdown examples being interpreted as real local references, hardened rulepack path containment and atomic adoption, refreshed bilingual documentation, and removed the CI wheel-version hardcode.
- Added isolated Scripted, Codex, and Claude Code host runners, normalized command/final-answer evidence, deterministic behavioral scoring, paired rollout statistics, feedback records, tournament selection, and explicit preliminary versus adoption-grade observed Gates.
- Kept static Skill-document replay as Gate 0 while separating it from observed Agent behavior and preserving stage-only Sleep and explicit human adoption.
- Completed the Evidence-Gated Evolution Lab through Phase 6: deterministic
inspect/summarize, Candidate and replay-task schemas, Evidence Only local cross-project collection, repository/date support counts, bounded rejection memory, a timeout-bound model adapter, and a localSKILL_EVOLUTIONrun history. - Hardened replay and Gate behavior: replay operates on temporary copies; validation and held-out suites must be byte-disjoint; Gate reports a quality vector plus token, command-surface, and workflow-overuse limits; Gate, stage, and adopt are hash-bound to the exact Patch IR; and a static no-network Gate Viewer supports human review.
- Bounded
aet learn sleepwith candidate, replay, model-call, and wall-clock limits, production-target change detection, and a stage-only terminal action. It never schedules, adopts, commits, pushes, uploads, or reads transcripts. - Rewrote English and Chinese README entrypoints around real user workflows and the actual architecture, including Hermes absorbed-Skill migration diagnosis.
- Added Evidence-Gated Evolution Lab (
aet learn): evidence-only harvest, deterministic failure-pattern mining, bounded rule or opt-in model Patch IR, isolated replay, immutable-contract/self-audit/held-out gates, stage, human adoption with a Decision Ledger entry, rejection memory, and a bounded localsleepcycle that can stage but never adopt, commit, push, or upload. - Added the Phase 0 evolution boundary, named editable blocks in the canonical Skill, core/validation/held-out/adversarial evaluation fixtures, and a metric vector acceptance policy rather than a synthetic trust score.
aet auditnow preserves a stale Hermes Skill reference as aFAILwhile detecting its local.absorbed_intomarker and emitting the installed replacement path in JSON/SARIF/Markdown remediation. This fixes the practical failure mode where a stale Skill Index was actionable only as a missing file.- Added regression coverage for the full rules proposal pipeline and the absorbed-Skill migration diagnostic.
- Added repeatable
aet trace --artifact <relative-path>for explicitly declared UTF-8 reports generated by a traced command, including pytest JUnit XML. AET redacts report content before persisting it and embeds the redacted artifact plus SHA-256 in the portable Evidence Pack. - Trace rejects absolute/outside-workspace artifact declarations. A missing,
non-regular, undecodable, or unredactable declared report is recorded as
UNKNOWN; a successful command then returns a non-zero Trace exit without rewriting the child command's successful execution result. Such a Trace also cannot advance a Run toPROVENor mark a bound proof as complete. - Added regression coverage for redacted report capture, portable-pack inclusion, missing artifacts, and outside-workspace rejection. This closes the evidence gap found while tracing Invest-Vault's pytest delivery proof.
- Constrained optional pytest discovery to AET's first-party
tests/directory so nested dogfood repositories are not accidentally collected into AET's own test report.
- Added an optional, deterministic Context Manifest:
aet context discoverrecords discoverable local instructions and Skills with SHA-256 hashes;recordadds local references and explicit read attestations;verifyreports changed/missing assets and workspace freshness. - Read declarations are deliberately stored as
agent_attestation(L5), while discovery and hashes remain L1 local evidence. AET does not claim that a model saw, understood, or used a recorded asset. - Added a local JSON Decision Ledger:
aet decision init,add,list,verify, andsupersedepreserve decision state, source hashes, evidence state, and replacement history. It is source-backed project memory, not RAG or a generic Agent-memory subsystem. - Added regression coverage for Context Manifest discovery/attestation/drift, Decision Ledger hash verification, and supersession; updated the portable Skill, bilingual README, package metadata, and wheel smoke target.
- Added an optional, append-only Run Manifest (
aet run init,status,verify, andclose) that describes a delivery lifecycle without becoming a workflow engine. Existingaudit,review,trace, andevidence packcommands remain independently usable and may opt in through--run. - Added declared lifecycle states:
INTENT_BOUND,AUDITED,REVIEWED,PROVEN,PACKED,STALE, andCLOSED. A Run records every transition and only closes a fresh, successfully packed evidence chain. - Extended
workspace_snapshotwith tracked-worktree, intent, and config fingerprints. Snapshot binding now distinguishesINTENT_CHANGED,CONFIG_CHANGED, andUNTRACKED_SET_CHANGEDfrom generic workspace or HEAD differences. - Added regression coverage for the full Run lifecycle, persisted stale state, intent/config changes, and changes to the untracked file set.
-
Added a shared
workspace_snapshotto audit, review, and Trace reports. It captures the Git HEAD plus deterministic tracked and untracked worktree digests without executing a declared proof command. -
Evidence Packs now compare every supplied artifact with the workspace at pack time through
snapshot_binding:EXACT_MATCH,HEAD_MATCH_WORKTREE_DIFFERS,HEAD_DIFFERS, or explicitUNKNOWN. A stale snapshot is reported separately from proof success; a command that passed is never rewritten as though it did not run. -
Reworked the static Evidence Viewer around delivery state, proof binding, and snapshot binding before exposing the raw JSON.
-
Added regression coverage for exact snapshot matches and changes made after a successful proof; refreshed the release workflow's wheel smoke target.
-
Reframed the README as a bilingual product entrypoint with an architecture, quality boundary, quick-start flows, Repo Archaeologist guide, and audience guidance.
-
Added a complete Simplified Chinese README, a contribution guide, a copyable intent example, and structured GitHub Issue forms.
- Added scoped
aet.tomlaudit policies with explicit exclusion reasons. - Added Evidence IR metadata, proof-bound Trace records, pack consistency checks, and an offline static Evidence Pack viewer.
- Added transparent
aet triage; it ranks work but never changes a finding'sPASS/FAIL/UNKNOWNstatus. - Added
aet evolve(Repo Archaeologist): plan, collect, build, report, and query stages; local Git/docs work offline and GitHub export/API use is explicit and evidence-manifested. - Added release governance, schemas, CI, a v1 Skill flow, regression fixtures, and release documentation.
- Added opt-in redacted command Trace and portable Evidence Pack compilation.