Skip to content

tasks.md

Latest commit

 

History

History
161 lines (153 loc) · 10.1 KB

tasks.md

File metadata and controls

161 lines (153 loc) · 10.1 KB
Raw

Tasks

Open

  • PluginService / runtime plugin lifecycle: the out-of-process plugin path (src/core/api/plugins/service.py, src/core/plugin/) is still outside the discovery transaction. It is omitted from the coverage kernel, has no equivalent rollback, and its PluginLoader.discover_plugins returns its own live mapping. Not addressed by the registry 1.4.0 work, open since 2026-08-08. This item is scoped to that runtime lifecycle only — it is not a statement about the plugin surface as a whole; other plugin surfaces are documented open separately (see the next item).

  • Other plugin surfaces documented open in this repository, listed so the item above is not read as covering them. Per docs/specs/PLUGIN_MANIFEST_SPEC.md: the flyto.plugin.v1 manifest is a DRAFT specification, not an implemented contract; RuntimeInvoker.set_plugin_manager has no caller anywhere, leaving _plugin_manager as None so a workflow step cannot reach a plugin subprocess; and the limits on what flyto-core can bound for a plugin running as its own process are recorded there rather than resolved. Neither the registry 1.4.0 work nor its acceptance touched or assessed these.

  • Add multi-route BFS crawler for arbitrary sites.

  • Connect Warroom evidence packs into Flyto2 Cloud UI.

  • Add enterprise airgap smoke recipes for no-egress browser/API checks.

  • Add recipe fixtures for capability states and report/export flows.

  • Keep maintained recipe bundle tests aligned with product-loop contracts.

  • Add restringer to reverse.deobfuscate once its npm publishing situation is resolved (currently maintained by an unofficial fork whose dependency tree dropped isolated-vm), or vendor it from HumanSecurity/restringer at a pinned commit instead.

  • Build a reliable ~/.flyto/node/ auto-download mechanism if a future need requires bundled (not system-installed) Node.js — reverse.deobfuscate sidesteps this by requiring a system Node.js instead of solving it.

Watch

  • Recipes that depend on local credentials or machine-specific paths.
  • Browser checks that pass without asserting real page state.
  • Module parameter drift that is not reflected in examples or tests.

Done

  • Closed the capability/extension/runtime change set on 2026-08-12: all pinned checks and 2,785 non-browser/e2e tests pass. Codex also fixed a routed-plugin policy alias that could omit the actual handler manifest's permissions; the regression pins both resolved identity and denial. Commit 0a353ff passed clean-tree strict Indexer verification 19/19.

  • Made plugin discovery a transaction over the registry, keyed on what each register_all() actually registered: stale modules a plugin stopped providing are removed, a failed load restores displaced rows to their real owner, and the contribution record is replayed only into a registry that began the pass empty. Closed the registry torn-read window in the same work: every public read holds _discovery_lock for its whole body, discover_plugins/refresh return a copied plugin mapping on every path, and PluginInfo is frozen so a caller cannot edit registry state through a value it was handed. REGISTRY_VERSION is 1.4.0. Verified and accepted 2026-08-11 on branch main: six pinned checks, Core module-contract proof, strict Indexer verification, and an independent 78 registry / 25 catalog test replay all passed. Accepted against these flyto coding implementation revisions (SHA-256, not Git commit hashes): docs — job_453f3754aa2041309060b75a / ebeb0ebfcab2d56bec576a944dcadd23fa197ff9726c558379df1c76eb12e341; source/tests — job_ad0baf4f580e4bc6aaac37de / b391189517db77146c4ab51def48ed7ada04fb30308296480e2e083df46bf65c; catalog/tests — job_8d8d49019afa402a8c503aa0 / a08df544401cf36a54dfe4f6fc084512cb3035a9febf885442baca5cd8366f15.

  • Settled the registry_plugin_contract coverage question: the pinned argv in .flyto/coding.yaml now passes --no-cov, so the check reports the registry contract it was pinned to prove rather than the project-wide coverage floor it inherited from addopts.

  • Hardened the remaining reported module boundaries: agent Ollama requests are localhost-only by default and use the guarded HTTP path, while the reported file readers and browser/document writers enforce FLYTO_SANDBOX_DIR before touching the filesystem. Updated supported dependency floors and CI tooling so the Python 3.10+ audited environment has no known package vulnerabilities.

  • Closed the reported outbound-request and cloud-download gaps: core API, OAuth2, and provider webhooks now use connect-time/per-redirect SSRF guards; Azure, GCS, and S3 downloads enforce the filesystem sandbox before writes.

  • Added reverse.deobfuscate (Phase 4 of the reverse.* toolkit): real semantic deobfuscation via webcrack, run in a dedicated Node.js sidecar worker spawned/killed per invocation, gated behind a new code.execute permission. Requires a system-installed Node.js 22/24. Reconciled the generated catalog to 468 modules / 85 categories.

  • Added reverse.request_breakpoint (set/remove/list request-level XHR/fetch breakpoints via CDP's DOMDebugger.setXHRBreakpoint) and session-snapshot reuse in reverse.attach (reattaching to the same still-enabled page returns the existing session's snapshot instead of detaching and rebuilding it, unless force_new=True). Reconciled the generated catalog to 467 modules / 85 categories.

  • Hardened the reverse.* toolkit: rewrote reverse.hook's injected JS on Object.defineProperty so hooks survive lazy (not-yet-defined) properties and later reassignment (narrower known limitation: immediate parent object must already exist), and added a shared session idle-timeout reaper (src/core/session_reaper.py, 30 min default, FLYTO_SESSION_IDLE_TIMEOUT_S override) wired into all three transports, reaping both browser_sessions and debugger_sessions uniformly. No catalog/module-count change — no new or changed module IDs.

  • Added reverse.sourcemap (resolve/list_sources/get_original_source): hand-rolled Source Map v3 VLQ decoder, no pip dependency, no permission gate, delegates external .map fetches to the already SSRF-guarded http.get instead of fetching anything itself. Strengthens Phase 1-3 rather than adding a new phase. Reconciled the generated catalog to 466 modules / 85 categories.

  • Added Phase 3 of the reverse.* toolkit: reverse.code (beautify/list_functions/list_strings/find_calls) — pure Python (tree-sitter + jsbeautifier, new jsast extra), no Node.js, no permission gate (never touches a browser/CDP session or executes code). Reconciled the generated catalog to 465 modules / 85 categories.

  • Added Phase 2 of the reverse.* toolkit: reverse.hook (install/remove/list/get_records), reverse.network (start/stop/list/get_initiator), and reverse.websocket (start/stop/list/get_frames) — all extending the same ReverseSession/CDP session Phase 1's reverse.attach creates. Reconciled the generated catalog to 464 modules / 85 categories.

  • Added the reverse.* CDP debugger module category (Phase 1 of the reverse-engineering toolkit): attach/detach, script list/get_source/search, breakpoint set/remove, wait_paused, resume, step, get_call_frames, and evaluate_on_call_frame, gated behind the new deny-by-default browser.debug permission. Reconciled the generated catalog to 461 modules / 85 categories.

  • Added the source-backed documentation contract, technical whitepaper, feature/API/configuration/security/operations/testing guides, generated declaration/module/CLI/HTTP/environment references, ownership manifest, and deterministic drift/link checks.

  • Added a Flyto2 brand and approved-mailbox policy gate for public repository content.

  • Declared crypto, dns, and ai package extras and aligned contributor dependencies with the tested capability set.

  • Protected workflow status and evidence reads with the Execution API bearer token and added regression coverage.

  • Expanded CI from one recipe test to documentation, brand, audited lint, non-browser tests, npm audit, package/Twine, and strict Indexer gates.

  • Added generic verification.* module IDs for deterministic verification primitives and kept warroom.* as compatibility aliases only; engine-owned workflows should compose verification.*.

  • Renamed the emitted deterministic testing model to the core-owned flyto.core.deterministic_verification.v1 while retaining warroom.automation_test_model.v1 as the legacy schema reference.

  • Kept Flyto2 product naming out of core defaults; engine artifacts may attach a product contract such as flyto2.automated_product_testing.v1.

  • Added explicit non-LLM execution policy to automation evidence: deterministic evidence is the fact source and gate authority; LLM is only an optional evidence reviewer.

  • Added event-stream and scheduler-loop summaries to automation_test_model.v1 so Product Verification evidence can distinguish callback/SSE contracts and durable scheduler contracts from raw replay facts.

  • Added automation_test_model.v1 to Warroom evidence packs so UI/CI can read coverage, scenario synthesis, replay, ghost API, invariant, RBAC, and evidence-chain results without interpreting raw JSON.

  • Added the 90-point Product Verification evidence gate to warroom.report with focused tests for pass and blocked paths.

  • Split warroom.run module success from replay_ok product success so failed deterministic replay evidence remains available to warroom.report.

  • Made generated Warroom DOM assertions hydration-aware so SPA pages are not scored as blank before body text settles.

  • Added a dedicated flyto-verification Docker image boundary for full-stack Product Verification compose/staging smoke.

  • Added operator-controlled SSRF dev-port allowlisting so Product Verification can replay local/staging browser targets without turning off the SSRF guard.

  • Added deterministic graph/scenario/report modules in flyto-core.

  • Made e2e/scenario test modules execute real module steps.

  • Bootstrapped project memory skeleton and lint gate.