CHANGELOG.md

Changelog

All notable changes to this project are documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

[Unreleased] — Sprint 10: v1.89.14 → v1.90.0 catch-up

Gap 1: Trajectory export — foundation + decryption + session sources (G10.1, partial)

• gep/trajectory/ (new package): ports the core of trajectoryExport.test.js.
  • builder.py — build_trajectories() / build_trajectory_from_rows(): group proxy-trace rows by session into evomap.coding_trajectory.v1 trajectories; per-turn extraction (provider, endpoint, response_id, previous_response_id, request/response bodies, reasoning, encrypted_content, per-turn tokens, error); tool-call extraction across Anthropic /v1/messages, OpenAI /v1/responses, /v1/chat/completions (declared tools vs actual invocations; Anthropic tool_use deduped by id); full streamed tool-argument reconstruction (Anthropic input_json_delta; OpenAI Chat delta + full-snapshot dedup; OpenAI Responses delta + .done override); Bedrock provider normalisation; language detection (keywords + file extensions); failure-correction marking; test-execution / code-edit / test_commands detection from tool inputs; stats (turns, tokens, has_tool_calls, tool_call_count, tool_types, has_test_execution, has_code_edit, test_commands).
  • io.py — write_trajectories(): atomic (temp + os.replace); a pre-placed symlink is not followed (PR #294 C4); owner-only 0o600 on POSIX.
  • crypto.py — read_trace_rows_detailed() / decrypt_trace_row(): AES-256-GCM under a node-secret-derived key with secret_version keyring selection; RSA-OAEP-SHA256 hub-key envelope unwrap with node-secret fallback; fail-closed (3 distinct messages) unless --allow-partial.
  • sources.py — non-proxy session logs: Codex rollout JSONL (session_meta + response_item records — message/reasoning/function_call/ function_call_output/custom_tool_call/tool_search*), Claude Code transcript JSONL (user/assistant with message.content blocks), and OpenAI generic-chat messages JSONL (top-level role-tagged records, prompt_tokens/completion_tokens, thinking+signature, OpenAI tool_calls/tool outputs) — with reasoning turns, custom tool calls, tool-search events, and test-execution / code-edit / failure-correction detection; plus detect_source() classification.
  • CLI: evolver trajectory --input <file|dir> --output [...] auto-detects session logs vs proxy traces, recurses directories, and decrypts with --node-secret/--hub-private-key/--node-secret-keyring/--allow-partial.
• tests/gep/trajectory/ (27 cases: 11 builder incl. full streaming + 10 crypto + 6 sources).
• Deferred (niche vendor sources): Cursor vscdb (SQLite), Gemini CLI+Gateway, Kimi Wire — bespoke low-frequency parsers, each with its own format-specific test file.

Gap 8: Force-update hardening (v1.90.0 contract)

• force_update.py (262→490+ lines): ports the portable subset of Node's forceUpdate*.test.js (Node-specific npx/degit/package.json/index.js/exit-78 mechanics are N/A for Python and intentionally omitted).
  • Sentinels — FORCE_UPDATE_BUSY / FORCE_UPDATE_NOOP (distinct singletons; identity-comparable, no truthy collision).
  • Concurrency guard — module-level mutex in execute_force_update(): a re-entrant call mid-upgrade returns FORCE_UPDATE_BUSY without re-downloading; mutex resets via finally (and on throw). (Fills a gap: the docstring claimed a file-lock guard that was never implemented.)
  • Idempotent floor — required_version is a minimum floor, not an exact target: operator (>=/>/=) + leading-v normalisation; an install that already satisfies the floor returns FORCE_UPDATE_NOOP (no downgrade, no re-download). Anti-downgrade guard (#213): an unparsable current version is refused, not silently satisfied.
  • Coded frozen failures — every failure is an immutable result carrying a stable code + detail; is_force_update_failure() + FORCE_UPDATE_FAIL_CODES registry.
  • Safe extraction — _safe_extract() refuses Zip-Slip (path-traversal) entries (keep-list/tarball-fallback safety).
  • report_force_update_outcome(noop/updated) persists status (skipped/ success); noop wins defensively.
• tests/test_force_update.py (+19 cases).

Gap 9: Outbound sync resilience (v1.90.0 contract)

• proxy/sync/outbound.py (108→290+ lines): ports proxyOutboundSync.test.js.
  • Body-size budgeting — one size-bounded batch per flush (EVOMAP_OUTBOUND_SYNC_MAX_BODY_BYTES env, overridable by store state after a 413); a single message that cannot fit is rejected, not sent.
  • 413 handling — single-message 413 quarantines; multi-message 413 backs the budget down and leaves all messages pending (1 Hub call).
  • Retryable vs terminal — retryable per-message failures defer (status pending, retry count untouched, next_retry_at set); terminal finalises. terminal wins over retry hints (PR #301).
  • proxy_trace gating — proxy_trace dropped when trace_collection_enabled store state is False.
  • Redaction — Hub non-2xx response text redacted before persistence.
  • Rich result shape: sent/synced/dropped/deferred/payload_too_large/ error/responses.
• proxy/mailbox/store.py: Message.next_retry_at field; poll_outbound skips deferred-not-due messages; new defer() (backoff without burning retry).
• tests/test_proxy_outbound_sync.py (new, 11 cases); test_proxy_sync.py updated to Node v1.90.0 sent=batch-size semantics.
• Encryption-envelope validation of proxy_trace payloads deferred to G10.1.

Gap 5: Host Error Classifier (#571)

• gep/host_error_classifier.py (new): is_host_client_error() + non-global HOST_PROVIDER_ERR_RE — classifies 4xx provider errors (invalid_api_key / insufficient_quota / rate limit / MaxTokens / HTTP 4xx) with bare-number-safe context. None/non-str/empty → False.
• gep/signals.py: under a host client error the failure-streak path is skipped — no ban_gene / failure_loop_detected / consecutive_failure_streak / force_innovation_after_repair_loop; the actionable host_llm_client_error signal is surfaced instead. An LLM quota/auth storm can no longer ban a gene.
• tests/gep/test_host_error_classifier.py (new, 5 cases): ports hostClientErrorSignals.test.js.

Gap 2: Solo mode (--solo / constrained-wild / "Mad Dog")

• solo/ subsystem (new): breaker.py (network "no escape valve" hard cut) + git_guard.py (local-git-only guard, wired into git_ops.run_cmd) + __init__.py (banner). Solo state = EVOLVER_SOLO env (process-wide, import-race-safe).
• cli.py: --solo flag (implies --loop); activates before dispatch so env overrides + hub cut land at the source. Even a user-set A2A_HUB_URL is ignored. Validator daemon + ATP auto-spend + task pickup are hard-cut in both the startup path (start_validator returns None; ATP envs forced off) and the in-cycle path (post_cycle guards).
• config.py: resolve_hub_url() returns "" under solo (no escape valve); new MAX_CYCLES_PER_PROCESS (EVOLVER_MAX_CYCLES_PER_PROCESS, 0=unlimited).
• evolve/runner.py: daemon loop honours MAX_CYCLES_PER_PROCESS (exits after N cycles — solo/CI testability).
• Fix: cli._cmd_loop now guards add_signal_handler/remove_signal_handler against NotImplementedError so --loop/--solo work on Windows (ProactorEventLoop lacks signal-handler support).
• tests/solo/test_solo.py (new, 11 cases): ports soloMode.test.js, including a subprocess smoke test asserting banner + service cut + clean exit.

Stats

• Tests: +73 (5 host-error + 11 solo + 11 outbound + 19 force-update + 27 trajectory); 0 regressions (1 pre-existing cognition test failure unrelated).
• Baseline: tracking v1.89.14 → v1.90.0 (G10.5, G10.2, G10.8, G10.9 closed; G10.1 trajectory core complete — proxy + full streaming + crypto + Codex/Claude/generic sources — Cursor/Gemini/Kimi vendor sources deferred; G10.3 cliContracts / G10.4 recipe pending).

[Unreleased] — Sprint 9: v1.89.14 parity (7 gaps closed)

Gap 1: Inert Gene Ban (#562)

• gep/memory_graph.py: stable_no_error/heuristic_delta/predictive outcomes now classified as inert — they build no Bayesian confidence and no longer count as successes for preferredGeneId.
• New _count_trailing_inert(): after GENE_INERT_BAN_STREAK (=8) consecutive trailing inert outcomes with no real success, the gene is added to bannedGeneIds so the selector yields null and the pipeline mutates.
• A single real success (e.g. error_cleared) resets the inert streak.
• 5 regression tests ported from test/issue562InertGeneBan.test.js.

Gap 2: Node Secret Versioning

• proxy/lifecycle/manager.py: parse_node_secret_version(), node_secret_version property (store > env precedence), stale-secret detection (store version < env version → Hub rotated → clear store secret).
• hello()/heartbeat() persist the Hub-returned node_secret_version.

Gap 3: Hub-Unreachable Exponential Backoff

• proxy/lifecycle/manager.py: _record_hub_unreachable() / _record_hub_reachable() / _hub_unreachable_wait_ms() / hub_unreachable_backoff_ms() — exponential backoff (5s→15min cap) on network errors (ConnectError/TimeoutException).
• hello()/heartbeat() check backoff before sending and record reachable/unreachable on success/failure.

Gap 4: Anti-Abuse Telemetry Heartbeat

• gep/anti_abuse_telemetry.py: build_heartbeat_anti_abuse() — privacy-preserving envelope with HMAC-pseudonymized device/workspace hashes, source-confidence labels (hub_required/hub_service/hub_observed), integrity hashes, task timing.
• config.py: ANTI_ABUSE_TELEMETRY_MODE (default heartbeat, explicit opt-out).
• proxy/lifecycle/manager.py: heartbeat meta.anti_abuse injection when mode=heartbeat.

Gap 5: Outcome Report Mode (P4-a Slice B)

• config.py: OUTCOME_REPORT_MODE (default off) + outcome_report_mode() resolver (on/enforce/true → on).

Gap 6: Force-Update from Heartbeat

• proxy/lifecycle/manager.py: _maybe_trigger_force_update_from_heartbeat() with EVOLVER_FORCE_UPDATE_RETRY_COOLDOWN_MS (default 5min) — prevents Hub from hot-spinning force-updates on every heartbeat.

Gap 7: Last-Update Ack

• proxy/lifecycle/manager.py: read_pending_last_update() / set_pending_last_update(); heartbeat carries last_update_ack + node_secret_version in payload.

Stats

• Tests: 1573 → 1609 (+36 new tests, 0 regressions)
• Baseline: tracking v1.89.11 → v1.89.14 parity on lifecycle + GEP selection

[Unreleased] — Sprint 0-8 catch-up against evolver v1.89.11

Sprint 0: Engineering baseline

• Fixed gep/sanitize.py import json position bug (was at file bottom, caused NameError).
• gep/sanitize.py: reverse leak scan now skips path/URL-shaped env values (#568).
• Added 6 credential redaction patterns (jwt, aws, github, slack, connection_string, high_entropy).
• CHANGELOG.md upgraded from stub to Keep-a-Changelog format.
• CONTRIBUTING.md: conventional commits with scope; baseline updated to 1331→1534 tests.

Sprint 1: IDE runtime hooks

• Rewrote adapters/scripts/runtime_paths.py (24→315 lines): host-env project dir resolution, workspace-id atomic create with symlink guards, FS-only fallback.
• Rewrote session_start.py (41→211): workspace-scoped memory recall, non-git notice (throttled), dedup, lazy memory read from newest end.
• Rewrote session_end.py (48→214): HEAD~1 diff, workspace-id stamping, Cursor systemMessage suppression.
• Rewrote signal_detect.py (39→160): context-aware stratification, Claude Code payload parsing, multilingual (CJK) fallback.
• New lock_paths.py (52 lines): daemon singleton-lock location + lease staleness tunables.
• New task_recall.py (103 lines): @evolver recall triggered capsule recall.
• Updated memory_filtering.py: added filter_relevant_outcomes (Node contract).

Sprint 3: Execution bridge + conversation sniffer

• New gep/exec_bridge.py (93 lines): Windows npm .cmd shim resolver (CVE-2024-27980).
• New gep/conversation_sniffer.py (240 lines): scan_corpus with local co-occurrence, off/shadow/enforce modes, cooldown, CJK support.
• Extended gep/bridge.py: added determine_bridge_enabled().
• evolve/runner.py: Ralph-loop stale bridge-mode break (#559).

Sprint 4: Security + seed library + deepening

• Expanded genes.seed.json (3→11 genes): full alignment with Node v1.87.0 seed library.
• Rewrote gep/skill2gep.py (187→400+): parse_skill_md with frontmatter + CJK sections, infer_category with word-boundary matching, skill_to_gene_dict with asset_id + quality heuristics.
• Deepened gep/idle_scheduler.py (220→300+): EVOLVER_IDLE_OVERRIDE, build-activity detection, FS-only idle fallback.
• gep/schemas/gene.py: added avoid and _source (alias) fields.

Sprint 5: Multi-provider proxy routes

• New proxy/router/gemini_route.py (160 lines): Google Gemini API proxy with SSE.
• New proxy/router/vertex_route.py (145 lines): Vertex AI proxy with ADC auth.
• New proxy/router/ollama_route.py (140 lines): local Ollama proxy.
• New proxy/router/responses_route.py (150 lines): OpenAI-compatible API proxy.
• New proxy/router/models_route.py (85 lines): /v1/models aggregator.
• New proxy/server/settings.py (62 lines): proxy settings persistence.
• New proxy/trace/extractor.py (65 lines): multi-format token usage extraction.
• New proxy/trace/usage.py (60 lines): usage aggregator singleton.
• New proxy/envelope.py (45 lines): structured message envelope.
• New proxy/inject.py (50 lines): context injection + internal field stripping.
• Updated model_router.py: 5-provider upstream detection.

Sprint 6: ATP CLI + mailbox transport

• Rewrote atp/cli.py (86→270): 15 subcommands (buy/orders/tasks/claim/deliver/settle/dispute/publish/policy/proofs/tier/order/status/enable/disable).
• Deepened atp/atp_execute.py (87→230): sandbox validation, structured proof building.
• Deepened atp/atp_task_pickup.py (99→200): ROI scoring, capability matching, concurrent limit.
• New gep/mailbox_transport.py (115 lines): proxy mailbox client with auto-start.

Sprint 7: Missing modules

• New gep/token_savings.py (120 lines): token/USD cost savings tracker with monthly reports.
• New gep/narrative_memory.py (85 lines): evolution history narrative compressor.
• New gep/memory_graph_adapter.py (120 lines): advanced queries (success trajectory, failure clustering, fuzzy match).
• New gep/directory_client.py (75 lines): EvoMap directory service client.
• New gep/oauth_login.py (115 lines): OAuth 2.0 device-code flow with keychain integration.
• New gep/claim_nudge.py (75 lines): throttled task-claim suggestion generator.
• New gep/device_id.py (85 lines): cross-platform anonymous hardware fingerprint.
• New gep/anti_abuse_telemetry.py (120 lines): abuse pattern detector (flood/bypass/exhaustion).

Sprint 8: Experiment framework + i18n + docs

• New experiment/ module (4 files): agent_runner, metrics, comparison, cli — controlled A/B evaluation.
• New README.ja-JP.md: Japanese README.
• New README.ko-KR.md: Korean README.

Stats

• Tests: 1331 → 1546+ (215+ new tests, 0 regressions)
• Source files: 192 → 217+ (25+ new modules)
• Seed genes: 3 → 11
• Proxy routes: 4 → 9
• ATP subcommands: 5 → 15
• mypy strict: 0 errors across all files
• Baseline comparison: v1.89.2 → tracking v1.89.11

[1.89.2] - 2026-06-09

• Initial Python port release tracking Node.js v1.89.2.
• GEP data layer, evolution pipeline, Proxy infrastructure, ATP marketplace (partial), IDE adapters (partial), WebUI (partial).