Download docs/V8.0_IMPLEMENTATION_PLAN.md from polyHedra/PolyHedra: direct link, hf CLI and curl.
- Browser
- Download file 67.7 kB
-
https://huggingface.co/spaces/polyHedra/PolyHedra/resolve/main/docs/V8.0_IMPLEMENTATION_PLAN.md
- Command line
-
hf download hf://spaces/polyHedra/PolyHedra/docs/V8.0_IMPLEMENTATION_PLAN.md
-
curl -L -o V8.0_IMPLEMENTATION_PLAN.md https://huggingface.co/spaces/polyHedra/PolyHedra/resolve/main/docs/V8.0_IMPLEMENTATION_PLAN.md
Polyhedra V8.0 Implementation Plan
Status: local implementation integrated; release verification is blocked on external live-provider stability and requested deployment checks. The shared coordinator, UI, and Slack entry paths invoke the governed typed runtime; all 19 agents have registered schemas/bindings, and VizAgent plus PortfolioMonitor use the same routing, ACL, trace, and run-record controls. Configured live UI runs have verified the provider roots, TrendAnalyst, MarketEventWatcher, and MacroContextWatcher. A full Finley run, general/delta/mixed live runs, live Slack, live support contexts, and a requested HF deployment remain open. The latest local Gemini verification encountered transient HTTP 503 and rate-limit responses before the remaining agents could run. Source of truth: polyhedraV8.0.md. Inherited behavior: V7.1 replay and evidence ledger, V7.0 session-delta YOLO, and the existing quant agents and Finley output.
Working Rules
- Keep the existing quant agent names and service boundaries. Implement inside the existing agents, core, services, messaging, and web modules; do not add version-named agent directories or a second product pipeline.
- Complete phases in order. Agent migrations may be developed incrementally, but release the governed quant runtime only after all registered quant agents pass the migration gate. Do not deploy an intermediate mixed workflow.
- Every function point below starts unchecked. Check it only after its code, relevant tests, and PRD acceptance criteria are verified. Record the evidence with the function-point ID in the Verification Record.
- Use existing provider configuration and dependencies. No new API keys, integrations, services, authentication tokens, or schedulers are authorized by this plan. Report a demonstrated provider entitlement gap before adding a replacement provider.
- Keep commits separated by function point or coherent module boundary. Commit and push when requested; deployment and live tests are separate from writing this plan.
- Preserve the halted distilled committee expansion and current trusted performance-weight behavior. This migration changes artifact contracts and execution control, not the experiment's analyst membership.
Phase 1: Contracts
Primary modules: agents/quant/outputs.py, agents/base.py, core/role_registry.py,
and a repo-packaged routing/ACL definition loaded by the existing runtime.
- C01 Encode all 19 routing entries, exact registered names, four scope enums, fixed asset symbols, execution contexts, and required dependency lists in one structured definition. Reject unknown agents and dependency cycles.
- C02 Encode the PRD filename ACL and baseline ACL in the same contract source. Ensure runtime loading works in the built package and deployment, without a dependency on desktop skills or developer machine paths.
- C03 Define common typed input, artifact, reference, evidence, data-issue, and delta models. Resolve every nested type named in the PRD to explicit fields, units, nullability, requiredness, and validation constraints.
- C03a Implement shared typed envelopes, JSON/rendering reference models, evidence locators, data issues, and common enums. Validate session/run identity, producer filenames, explicit timezone, unique scope, and schema round trips. C03 remains open until all agent-specific nested types are resolved.
- C04 Implement the DataFetcher manifest and per-ticker MarketDataArtifact models, including component applicability, empty successful feeds, partial coverage, freshness, and provider errors without ticker status suffixes.
- C05 Implement typed input/output models for each remaining agent, preserving distinct concrete producer literals for the four asset analysts.
- C05a Implement OptionsFlowWatcher input/output and nested options schemas. Enforce options scope, bounded sampling, source capability flags, exact candidate rules, non-directional candidates, and evidence-backed unusual flow.
- C05b Implement TrendAnalyst, SentimentAnalyst, MarketEventWatcher, MacroContextWatcher, UnusualEventDetector, and EventRouter input/output schemas. Enforce scope-filtered dependencies, evidence ownership, explicit unavailable feeds, conditional options overlap, and complete event assignment accounting.
- C05c Implement distinct BTCAnalyst, SPYAnalyst, QQQAnalyst, and GLDAnalyst input/output schemas plus the runtime-owned AssetAnalysisCollection. Enforce fixed ticker identity, evidence-backed signals, conditional paper-trade fields, and explicit unavailable contributions.
- C05d Implement QuantResearcher, Backtester, RiskAnalyst, and QuantStrategist input/output schemas. Separate historical simulation from LLM assessment, preserve active general/specialized joins, and keep committee performance fields outside the LLM-owned payload.
- C05e Implement Publisher, VizAgent, and PortfolioMonitor input/output schemas. Enforce Finley's explicit no-proposal reason, chart source/manifest identity, and complete paper-position monitoring without close authority.
- C06 Define reference identity and schema-version validation for JSON and rendering references. Distinguish LONG/SHORT trade direction from analytical market direction. Keep unknown values null rather than zero or fake metrics.
- C07 Validate conditional decisions: trade levels only for actual paper suggestions, no conflicting symbols, at most three final proposals, rejected signals excluded, and elevated risk accompanied by mitigation.
- C08 Keep data availability separate from execution state. A recorded failure must not validate as a successful market NO_TRADE conclusion.
- C09 Add a provider-neutral structured completion boundary using existing adapters: strict closed-schema validation, one JSON object, at most one format correction, explicit validation errors, and usage totals for completed attempts. Add a typed input/prompt/output/Markdown binding primitive with validation before prompt construction and rendering. Verify this boundary with controlled responses.
- C10 Register concrete input schema, prompt template, output payload schema, and Markdown renderer for every quant agent. Verify schema coverage and agent-specific renderings; use these bindings on the actual migrated paths in M01-M19.
- C10a Add an immutable 19-agent binding registry with exact input models, LLM-owned output schemas, role-specific prompts, and deterministic Markdown renderers. Keep C10 open until every migrated runtime path uses its binding.
Gate: positive and negative schema cases, registry coverage, filename ownership, scope enums, and cycle validation pass. Verify serialized round trips and malformed LLM/provider payloads, not just class construction.
Phase 2: Session Artifacts And ACL
Primary modules: services/file_writer.py, core/session.py,
core/yolo_context.py, existing RAG/context services, and run persistence in app.py.
- A01 Allocate unique run IDs for every execution, including same-day runs; create the PRD run directory and retain the existing session identity.
- A02 Inject a scoped file interface enforcing read/create/edit/list, path resolution, ticker scope, current/source/baseline run bindings, and read-only published outputs. Block traversal, cross-session paths, and symlink escapes through the same interface.
- A02a Implement the scoped FileWriter view and shared publication registry. Resolve the packaged ACL against frozen run/ticker bindings, separate VizAgent source reads from output writes, and reserve performance writes for its trusted tracker capability. Keep A02 open until the migrated coordinator injects it.
- A03 Apply permissions to provider writes, renderers, prompt construction, and RAG retrieval. Ensure no helper bypasses the injected file boundary.
- A03a Apply the resolved read ACL to quant RAG retrieval before semantic ranking and defensively recheck returned artifact paths and ticker metadata. Keep A03 open until every migrated provider, prompt builder, renderer, and agent RAG call receives its scoped runtime interface.
- A04 Validate and publish JSON atomically; record actual typed input snapshots and artifact references. Preserve valid analysis when rendering fails, with the rendering error recorded separately.
- A04a Implement strict typed input snapshots and JSON-first atomic text publication through the scoped writer. Run the fixed Markdown renderer only after immutable JSON publication, and return rendering failures separately. Keep A04 open until migrated agents record references and binary chart publication is atomic.
- A05 Enforce one owner per agent output. Build the runtime-owned AssetAnalysisCollection from the selected run's analyst outputs, with explicit missing/unavailable contributions and scope filtering for each consumer.
- A05a Implement runtime assembly from immutable published specialized outputs, preserving a published unavailable analyst result separately from a missing result and returning an immutable collection filtered to each consumer's scope. Keep A05 open until dependency execution publishes and injects the aggregate on migrated runs.
- A06 Select and freeze an eligible prior completed YOLO run for delta context. Handle new tickers, unavailable baseline fields, and unsupported schema versions explicitly. Failed runs do not replace the eligible baseline.
- A06a Implement immutable typed-baseline selection over V8 run-record projections. Select the latest eligible completed YOLO run, retain an older eligible baseline when a newer run failed, report new tickers, and record explicit rejection reasons for unsupported, cross-session, support-context, current, and invalid records. Keep A06 open until V8 run records invoke the selector and persist the frozen baseline ID before activation.
- A07 Preserve legacy replay/files and session paper positions. Establish a typed baseline when only legacy artifacts exist; do not parse old Markdown into invented typed evidence or overwrite historical snapshots.
- A08 Extend the existing export to include all run artifacts, inputs, renderings, charts, references, replay records, and available evidence ledger. Verify all references inside a downloaded multi-run session resolve correctly.
Gate: isolated session tests demonstrate ownership, immutable publication, baseline selection, RAG filtering, same-day run isolation, export completeness, and preservation of existing files. No runtime producer has blanket file access.
Phase 3: Dependency Execution
Primary modules: agents/coordinator.py, core/agent_router.py,
core/eventbus.py, core/template_builder.py, core/cost_gate.py, and
core/state_patcher.py.
- R01 Build the active routing plan once from execution context and normalized tickers; store supported/effective scope and pruned dependency lists.
- R01a Implement the immutable active-plan builder from the packaged 19-agent Routing Table. Normalize and deduplicate requested tickers, preserve declared and effective scope, attach overlap scope to each applicable dependency, and retain explicit inactive reasons for registry inspection. Keep R01 open until the coordinator creates and persists one plan per run.
- R02 Activate ready agents in parallel and publish each completion once. Evaluate overlapping scope at joins; a BTC run never waits for ETF options.
- R02a Implement one-shot asynchronous activation from a frozen plan. Start all ready agents concurrently, wait on overlap-applicable dependencies, and reject duplicate or activation-mismatched publication identities. Keep R02 open until migrated agent execution publishes through this executor.
- R03 Resolve inputs from published dependency artifacts and the ACL. Capture optional inputs at activation; do not race-read mutable upstream files.
- R03a Register each agent's complete published artifact envelope and add an ACL-bound resolver that freezes required, optional, and selected-baseline JSON identities and contents at capture time. Reject unpublished, unreadable, cross-run, schema-invalid, or reference-mismatched artifacts. Keep R03 open until the executor captures these snapshots and migrated agents build their exact typed inputs from them.
- R03b Add a scheduler capture hook that resolves inputs synchronously after all applicable dependencies complete and before the agent task is scheduled. Pass the frozen value to the runner; record capture errors as the named agent failure while unrelated branches continue. Keep R03 open until product integration supplies the ACL-bound resolver and typed input builders.
- R04 Propagate typed partial/unavailable results without treating missing outputs as success. Record failed agents and the blocked dependents by name; allow unrelated ready work to finish under the same fixed plan.
- R04a Derive completed, failed, and blocked records in the dependency executor. Treat published COMPLETE, PARTIAL, and UNAVAILABLE artifacts as completed executions; block failed dependents transitively while allowing unrelated running and ready agents to finish. Keep R04 open until persisted V8 run records and the product UI consume these states.
- R05 Remove next-role authority from agent prompts, response mentions, cycle guards, and deterministic branch helpers on governed quant execution. EventRouter assigns evidence only and cannot expand scope or activate agents.
- R06 Preserve cost/token gates, trace logging, and coordinated state patches during parallel work. Prevent duplicate strategy/performance updates.
- R07 Keep human outside the V8 activation table and automated run. Map
legacy ROUTE_TO_HUMAN labels to
ANALYSIS_UNAVAILABLE; approval metadata must not create a human checkpoint, route, or automatic trading action in V8. - R08 Derive run and agent statuses from activation/publication/error records. A quiet LLM request is not blocked solely because Slack is silent.
- R08a Implement immutable typed terminal run records derived from the frozen routing plan and dependency execution result. Persist baseline mode, input snapshot paths, and the artifact inventory atomically beside legacy replay files. Keep R08 open until coordinator lifecycle updates and the UI consume these records.
Gate: controlled async tests prove overlap, join ordering, one activation per agent, mixed ticker scopes, failed dependency handling, and deterministic input selection. Use controlled producers for scheduler tests; these are not live LLM or provider verification.
Phase 4: Agent Migration
For each checkbox, migrate input construction, real agent execution, structured parsing, artifact publication, baseline context, and declared output rendering. Verify the agent's exact input/output model and ACL using representative data and an explicit missing-input case. Keep existing reasoning duties intact.
- M01 DataFetcher: retain fetch/normalization and computed indicators; persist component quality and actual timestamps through its ACL.
- M01a Add typed execution to the existing DataFetcher class. Reuse the configured Finnhub context and Yahoo daily-price path without provider-side V8 writes; normalize raw values, preserve actual OHLCV source provenance, publish one immutable typed artifact per ticker plus the manifest, and represent provider loss as typed availability. Keep M01 open until coordinator integration proves this entry point is the active product path. Approved 2026-10-24 provider scope: use Yahoo daily OHLCV; report an explicit data gap if it has no valid bars. Keep Finnhub quote/news/earnings, but do not call or smoke-test the inaccessible Finnhub candle endpoint.
- M02 OptionsFlowWatcher: preserve normalized contracts and bounded sampling; distinguish market context, candidates, and evidenced unusual flow. Retain missing versus zero values and record unsupported flow capabilities.
- M02a Add typed execution to the existing OptionsFlowWatcher class. Reuse the bounded yfinance provider, preserve sampled contracts and provider capabilities, classify threshold matches only as non-directional candidates, and publish provider loss as typed unavailable evidence. Keep M02 open until coordinator integration and baseline delta comparison are verified.
- M03 TrendAnalyst: perform LLM technical interpretation from valid typed prices/indicators and publish cited trend records with explicit data gaps.
- M03a Add typed execution to the existing TrendAnalyst class. Use the registered structured LLM prompt and payload, verify citations against frozen current and baseline inputs, preserve source quality issues, compare approved prior trend context in delta mode, and publish explicit unavailable analysis without calling the LLM when no valid OHLCV evidence exists. Keep M03 open until coordinator, cost/trace, and live configured-LLM integration are verified.
- M04 SentimentAnalyst: perform LLM sentiment interpretation for GENERAL scope; distinguish unscored or thin news from genuinely neutral evidence.
- M04a Add typed execution to the existing SentimentAnalyst class. Use the registered structured LLM prompt and payload, verify current and baseline citations against frozen inputs, preserve exact scored/unscored feed counts and source quality issues, compare approved prior sentiment in delta mode, and publish explicit unavailable analysis without calling the LLM when news and earnings evidence are absent. Keep M04 open until coordinator, cost/trace, and live configured-LLM integration are verified.
- M05 MarketEventWatcher: organize DataFetcher evidence and compare prior context without asserting a nonexistent live event provider.
- M05a Add typed execution to the existing MarketEventWatcher class. Use the registered structured LLM prompt and payload, constrain observed changes and catalysts to frozen current DataFetcher evidence, verify copied catalyst fields against cited news records, require cited approved prior context for delta claims, and expose the absent external event feed as an explicit typed gap. Keep M05 open until coordinator, cost/trace, and live configured-LLM integration are verified.
- M06 MacroContextWatcher: preserve LLM main-line interpretation, weighted drivers, scope, contradictions, and uncertainty.
- M06a Add typed execution to the existing MacroContextWatcher class. Use
the registered structured LLM prompt and payload, resolve the conditional
current watcher dependencies and approved prior macro context, require every
driver to have positive weight and current cited evidence, preserve upstream
gaps and exact affected scope, and publish deterministic
UNCLEARcontext without LLM inference when no usable current evidence exists. Keep M06 open until coordinator, cost/trace, and live configured-LLM integration are verified. - M07 UnusualEventDetector: preserve LLM event assessment and record meaningful, suppressed, unchanged, and unsupported evidence distinctly.
- M07a Add typed execution to the existing UnusualEventDetector class. Use the registered structured LLM prompt and payload, resolve conditional current watcher dependencies and approved prior events, require stable globally unique IDs and current evidence for meaningful events, verify unchanged and carried delta events against cited prior context, and publish deterministic monitoring conditions without LLM inference when current evidence is absent. Keep M07 open until coordinator, cost/trace, and live configured-LLM integration are verified.
- M08 EventRouter: publish typed event-to-ticker assignments within the initialized scope; never bypass the fixed routing plan.
- M08a Add typed execution to the existing EventRouter class. Use the registered structured LLM prompt only when meaningful events exist, account for every meaningful event as assigned or explicitly unassigned, require positive relevance and an exact citation to the current event record, and reject assignments outside the event's affected-ticker set. Publish an empty route without LLM inference when there are no meaningful events. EventRouter does not select agents, emit next roles, change activation, or expand scope. Keep M08 open until coordinator, cost/trace, and live configured-LLM integration are verified.
- M09 BTCAnalyst: migrate its BTC-specific reasoning and delta contract.
- M09a Add typed execution to the existing BTCAnalyst class with its fixed BTC schema, crypto-specific registered prompt, current evidence verification, read-only baseline resolution, delta citation, data-gap preservation, and deterministic unavailable output. Keep M09 open until coordinator, cost/trace, and live configured-LLM integration are verified.
- M10 SPYAnalyst: migrate its SPY-specific reasoning and delta contract.
- M10a Add typed execution to the existing SPYAnalyst class with its fixed SPY schema, broad-market-specific registered prompt, exact routed-event citation, current evidence verification, read-only baseline resolution, delta citation, data-gap preservation, and deterministic unavailable output. Keep M10 open until coordinator, cost/trace, and live configured-LLM integration are verified.
- M11 QQQAnalyst: migrate its QQQ-specific reasoning and delta contract.
- M11a Add typed execution to the existing QQQAnalyst class with its fixed QQQ schema, technology/growth-specific registered prompt, current evidence verification, read-only baseline resolution, delta citation, data-gap preservation, and deterministic unavailable output. Keep M11 open until coordinator, cost/trace, and live configured-LLM integration are verified.
- M12 GLDAnalyst: migrate its GLD-specific reasoning and delta contract.
- M12a Add typed execution to the existing GLDAnalyst class with its fixed GLD schema, gold/rates-specific registered prompt, current evidence verification, read-only baseline resolution, delta citation, data-gap preservation, and deterministic unavailable output. Keep M12 open until coordinator, cost/trace, and live configured-LLM integration are verified.
- M13 QuantResearcher: consume the conditional join, retain signal IDs across delta updates, and allow an evidence-supported empty signal result.
- M13a Add typed execution to the existing QuantResearcher class. Resolve and verify the active general and/or runtime-owned specialized branches, preserve upstream gaps, permit an LLM-owned empty signal list when current evidence is usable, publish explicit unavailable research without LLM inference when it is not, and require carried delta signals to retain their IDs and cite the prior signal record. Keep M13 open until coordinator, cost/trace, and live configured-LLM integration are verified.
- M14 Backtester: retain validation duties and distinguish LLM assessment from executed historical simulation. Unsupported metrics remain null with reasons; this phase does not introduce a historical execution engine.
- M14a Add typed execution to the existing Backtester class. Resolve and
verify current signals, optional market data, and approved read-only baseline
artifacts; require one exact cited result per current signal; publish only
LLM_ASSESSMENTwith null metrics and explicit reasons because no historical execution engine is approved; and skip LLM inference for empty or unavailable signal sets. Keep M14 open until coordinator, cost/trace, and live configured-LLM integration are verified. - M15 RiskAnalyst: retain signal verdicts and paper-portfolio context; forward a completed risk result even when all candidate signals are rejected.
- M15a Add typed execution to the existing RiskAnalyst class. Resolve and verify the current backtest and optional analysis context, consume explicit scope-filtered paper-position state, require one exactly cited verdict per backtest result, reject untested signals with zero exposure, validate deterministic portfolio heat, preserve upstream gaps, and publish completed all-rejected and portfolio-only assessments. Keep M15 open until coordinator, cost/trace, and live configured-LLM integration are verified.
- M16 QuantStrategist: retain proposal/no-trade/watch reasoning, lifecycle updates, committee votes, and trusted weights. Do not infer a decision from fallback error prose or permit the LLM to alter member weights.
- M16a Add typed execution to the existing QuantStrategist class. Resolve and verify current and approved baseline inputs, retain the existing four committee lenses, require current cited votes and complete signal/lifecycle accounting, validate proposals against current risk, attach trusted weights and deterministic weighted aggregation after LLM validation, preserve gaps, and publish unavailable strategy without LLM inference when risk is unavailable. Keep M16 open until trusted performance-state persistence, coordinator, cost/trace, and live configured-LLM integration are verified.
- M17 Publisher: retain Finley's readable baseline/delta output and exact proposal values; distinguish data loss, unavailable analysis, genuine no-trade, and watch outcomes without a required human activation.
- M17a Add typed execution to the existing Publisher class. Resolve and verify current strategy, optional context, and approved prior Finley output; require the LLM to produce the cited interpretation and reason classification; render proposal identifiers, levels, conditions, risk fields, and ordering deterministically from the validated strategy artifact; preserve upstream gaps; and publish the final Markdown body under 500 words with the paper-trade disclaimer last. Keep M17 open until the coordinator and Slack product path invoke this entry point with live configured-LLM verification.
- M18 VizAgent: use typed chart sources, persist real PNG bytes through the allowed interface, and record chart sources; remain outside normal YOLO.
- M18a Add typed execution to the existing VizAgent class. Resolve only explicitly selected, published JSON sources through the visualization ACL; render OHLCV, indicator, executed-backtest, risk, and monitored-position fields without Markdown parsing; publish atomic PNG bytes with manifest-bound references; use the LLM only for a structured caption after deterministic rendering; and publish explicit unavailable chart errors when required typed fields are absent. Keep M18 open until the on-demand product coordinator and UI invoke this entry point with a configured LLM and deployed renderer.
- M19 PortfolioMonitor: use typed open-position/quote inputs, preserve threshold checks and provenance, and publish alerts without closing positions.
- M19a Add typed execution to the existing PortfolioMonitor class. Consume runtime-supplied paper positions, thresholds, pause state, market-hours state, and existing quote-provider configuration; fetch each ticker once through an injectable existing-provider boundary; calculate direction-aware moves and threshold states deterministically; account for every input position; publish quote provenance, explicit gaps, and typed alerts; and reject any output that changes immutable position fields or threshold math. Keep M19 open until the separately initialized monitor coordinator, state timestamp, and alert product delivery invoke this path with live quote verification.
Gate: all 19 migrations pass agent-focused tests and integrated typed handoff tests. A model existing in outputs.py is not sufficient: runtime traces and persisted artifacts must demonstrate use on the actual execution path.
Phase 5: Runtime And Product Integration
- I01 Make UI and Slack invoke the same coordinator, registry, routing plan, session state, provider configuration, and typed artifact runtime.
- I02 Preserve existing quant entry commands and session behavior. Scope determines specialization; avoid a new version-specific product command.
- I03 Remove academic-specific agents, registrations, prompts, commands, and product descriptions from the quant-only runtime. Trace shared imports before removal; preserve quant RAG, LLM, session, and export infrastructure.
- I04 Persist run-specific routing and permission snapshots in the existing replay record. Keep recorded failures and last available Finley output associated with the correct run, without reusing another run's result.
- I05 Build evidence ledgers from actual input/output references and recorded exclusions. Do not infer use simply because a file exists.
- I06 Keep the session's trusted strategist performance state and each run's recorded weights consistent through delta runs and replay.
- I07 Update README and quant help to describe the latest whole product, preserving inherited capabilities and distinguishing deployed behavior from planned V8 changes. Keep historical PRDs intact.
Gate: regression tests cover session creation/resume, YOLO entry, general and specialized markets, traces, RAG, replay, package export, proposal state, and support contexts without academic imports or a parallel V8 workflow.
Phase 6: UI
Primary modules: web/static/index.html, existing session/run payloads and
routes in app.py, and tests/test_v611_web_ui.py.
- U01 Retain the YOLO entry, session/run selection, Finley output, options context, and download behavior; add Run, Routing Table, and Permissions views.
- U02 Render all running agents and dependency-based waiting states from runtime records. Show result availability separately from execution status.
- U03 Show the active routing plan by default and an explicit full registry view with inactivity reasons. Keep exact names, list dependencies, and scope enums.
- U04 Display per-agent ACL templates and resolved paths, input/output references, baseline source, and published read-only state. Keep views read-only.
- U05 Show genuine no-trade/watch decisions and data/agent failures distinctly; do not render a blocked dependency or pending analysis as a final market opinion.
- U06 Keep polling tied to the selected session/run. Preserve tab, selection, and scroll; a live run cannot overwrite a prior replay view. Label unavailable legacy routing/ACL information without filling it from current configuration.
- U07 Verify loading, empty session, failed refresh, absent replay, unavailable options, partial results, and download in every terminal state.
- U08 Verify browser screenshots and interactions at 390 x 844 and 1440 x 900: readable long paths, usable tables, keyboard controls, parallel rows, and no overlap or hidden result text.
Gate: backend payload tests and browser interaction checks pass. A static HTML preview alone does not verify runtime progress or replay behavior.
Phase 7: Release Verification
- V01 Run focused suites during each phase, then the full applicable quant, provider, coordinator, RAG, session, Slack, and UI regression suites once the migration is integrated. Record pre-existing unrelated failures separately.
- V02 Verify a build/package includes contracts and UI assets and starts with the existing deployment configuration. Do not claim configured means authenticated or provider-entitled until live requests verify it.
- V03 Verify live Slack YOLO for SPY and a general ticker through the same migrated agent runtime: real data, real LLM reasoning where required, trace usage, named failures when any, and useful Finley output in Slack.
- V04 Verify live UI baseline then delta in the same session, including BTC options exclusion and a mixed general/specialized run. Confirm the observed plan, parallel execution, and frozen tickers match the contract.
- V05 Verify two same-day runs, replay of the first, and downloaded inputs, outputs, evidence ledger, and baseline references against each run record.
- V06 Verify a controlled unavailable-input or failed-agent case through the product integration test setup. Do not deliberately break shared live credentials to create a failure; distinguish this test from live success evidence.
- V07 After a requested HF deployment, record deployed commit, startup status, live UI/Slack result, provider coverage/limitations, and download integrity. An actual unmet input requirement remains a release blocker.
Release gate: every required function point has evidence at its intended level. Mocked agents prove orchestration behavior only. Real LLM calls alone do not prove valid data, correct reasoning, usable results, or complete integration.
Commit Boundaries
Prefer coherent commits for contracts, file/ACL enforcement, dependency execution, fetch/context agents, specialized analysts, research/risk/strategy, Publisher/support agents, product integration, UI, and verification/documentation. Keep intermediate commits reviewable and avoid deploying partial migration.
Verification Record
Add entries as work is completed. Include function-point IDs, commit, actual check, outcome, and remaining limitation; live checks also include run/session IDs and deployment revision. Entries below distinguish contract tests from actual migrated runtime or live verification.
| Function points | Commit | Verification and level | Result | Limitation or follow-up |
|---|---|---|---|---|
| C01, C02 | 4bebf98 |
tests/test_quant_contracts.py: exact parity with all 19 PRD routing rows, current/baseline ACL tables, JSON round trip, invalid graph/identity/ownership/path cases. Built wheel with declared Hatchling backend and imported all 19 contracts directly from it. |
Passed, 18 tests plus wheel import | Declarations only; enforcement and activation remain A02/R01-R03. No deployment. |
| C09 | f816c35 |
tests/test_services/test_structured_completion.py and tests/test_agent_io_template.py: controlled typed input -> LLM -> validated payload -> fixed Markdown; strict types, unknown fields, malformed/duplicate/non-finite JSON, correction bounds, usage, provider errors, and revalidation. Existing test_llm_service.py regression suite included. |
Passed; 73 tests across the four focused files | No live provider requests or native constrained decoding. Concrete agent bindings, cost-gate integration, publication, and migration remain pending. |
| C03a | 32e44f2 |
tests/test_artifacts.py: positive/negative envelope, reference, evidence, rendering, scope, and enum tests. |
Passed locally | Generic foundation only. Resolving files, checking evidence existence, effective-scope filtering, and per-agent payload validation remain pending. |
| C04 | 5531cf2 |
tests/test_market_data_contracts.py: invalid/missing OHLCV, numeric types, zero versus unknown values, indicator warm-up, provenance/counts, empty feeds, applicability, freshness, baseline identity, and manifest coverage. Contract, LLM service, and Finnhub suites totaled 151 passing tests. An expanded quant run also had 164 passes and one pre-existing failure. |
Contract tests passed | test_quantstrategist_allows_zero_trade_proposals expects an empty state patch despite committee bookkeeping; reproduced unchanged against archived commit 448f145. M01 provider/agent migration and live verification remain open. |
| C05a | ebeaf4b |
tests/test_options_contracts.py: options scope/cutoff, prior-run ACL, strict contract values, sampling limits, exact candidate criteria, non-directional candidates, provider capability enforcement, and unavailable-data cases. Combined focused contract and existing options tests: 182 passed. |
Contract tests passed | M02 provider/agent migration and live verification remain open. |
| C05b | 8b01e73 |
tests/test_analysis_contracts.py: fixed fields, complete scope coverage, current-run dependency identity, conditional options dependencies, insufficient sentiment evidence, unavailable external event feed, macro evidence, meaningful/suppressed event separation, and complete in-scope assignment. |
Contract tests passed | M03-M08 agent migration and live verification remain open. |
| C05c | ba9d2fe |
tests/test_asset_contracts.py: distinct producer/ticker identities, current dependency validation, BTC options exclusion, evidence-backed decisions, conditional and ordered trade fields, explicit unavailable state, and runtime collection integrity. |
Contract tests passed | M09-M12 agent migration and live verification remain open. |
| C05d | 8dc524a |
tests/test_decision_contracts.py: active general/specialized joins, stable signal identity, empty-signal outcome, historical versus LLM assessment, coherent metrics, risk mitigation/rejection, conditional proposals, proposal cap, runtime-owned committee weights, and current dependency identity. |
Contract tests passed; 12 tests | M13-M16 agent migration, automatic risk cross-validation, trusted performance-tool adaptation, and live verification remain open. |
| C03, C05-C08 | 0733d82, 652c4ed |
tests/test_product_contracts.py plus the complete contract suite: Finley proposal/reason consistency and no routing field, chart source/manifest identity, unavailable chart behavior, monitoring scope/quote state/position accounting, shared unavailable-artifact gaps, and unavailable strategy separation from NO_TRADE. |
Passed; 164 tests across eight contract files | C10 bindings and all runtime migration function points remain open. No live provider or deployment verification. |
| C10a | 1a70cc9 |
tests/test_quant_io_registry.py and tests/test_agent_io_template.py: exact 19-agent coverage, concrete JSON schemas, unique role prompts, typed prompt construction, distinct specialized payloads, exclusion of runtime committee fields, fixed Markdown, Slack body rendering, and immutable lookup. |
Passed; 14 tests | Registry is not yet invoked by existing agents; C10 and M01-M19 remain open. |
| A01 | 9cd328c |
tests/test_yolo_context.py and tests/test_v611_web_ui.py: shared Slack/UI YOLO preparation now allocates a UUID-suffixed UTC run ID, creates quant/runs/{run_id}, persists the identity in QuantState, and propagates it through coordinator metadata and UI run records. Two runs at the same timestamp receive distinct directories. |
Passed; 44 tests | Support-context allocation and run manifests will be added with their owning migrations; no agent writes to the new directory until A02-A05/M01-M19. |
| A02a | 4b33011 |
tests/test_scoped_file_writer.py: owner and ticker enforcement, shared publication immutability, explicit read-only baseline binding, fixed VizAgent source run, traversal/absolute/backslash/wildcard rejection, declared-only listing, and trusted performance capability. |
Passed; 12 tests, 1 symlink test skipped because directory symlink creation is unavailable on this Windows host | Coordinator/provider/renderer/RAG injection remains A02/A03. Atomic typed publication remains A04. Existing FileWriter behavior is unchanged. |
| A04a | 158fc2b |
tests/test_artifact_publisher.py plus scoped-writer and I/O-registry regressions: exact typed input snapshots, identity checks before writes, atomic replacement failure preservation, JSON-first immutable publication, fixed Markdown rendering, and separate rendering-error capture. The complete contract regression suite passed 195 tests. |
Passed; 195 tests, 1 symlink test skipped because directory symlink creation is unavailable on this Windows host | Migrated agent paths do not yet consume the publisher or record its references. Binary chart publication and full A04 remain open. |
| A05a | 4f0c027 |
tests/test_asset_analysis_collection.py plus asset, decision, product, and complete typed-contract regressions: only registered published specialized outputs are accepted; missing and published-unavailable results remain distinct; producer/run/session identity, duplicate rejection, explicit gaps, stable source references, and consumer-scope filtering are enforced. |
Passed; 200 tests, 1 symlink test skipped because directory symlink creation is unavailable on this Windows host | Runtime publication/injection remains A05/R03 and agent migration. The collection adds no analysis of its own. |
| A06a | c690519 |
tests/test_baseline_selector.py plus artifact, legacy YOLO-context, and scoped-writer regressions: latest eligible completion, failed-run fallback, cross-session/current/support-run exclusion, unsupported schema, invalid inventory, new-ticker reporting, and immutable selection were exercised. |
Passed; 70 tests, 1 symlink test skipped because directory symlink creation is unavailable on this Windows host | The selector is intentionally dormant until typed V8 run records are persisted; existing V7 date-named baseline behavior is unchanged during migration. A06/R01/R08 remain open. |
| R01a | 65b535c |
tests/test_active_routing.py and tests/test_quant_contracts.py: general-only, SPY specialized, BTC without options, mixed general/specialized, visualization, and portfolio-monitor plans; normalized scope, fixed analyst symbols, overlap-aware dependencies, full 19-agent accounting, invalid scope, and immutability. |
Passed; 24 focused tests. The expanded routing run had 35 passes and the documented pre-existing QuantStrategist empty-state-patch failure. | Builder is not yet invoked or persisted by the coordinator; R01 and runtime migration remain open. No LLM routing behavior changed. |
| R02a, R04a | ea2c230 |
tests/test_dependency_executor.py, tests/test_active_routing.py, and tests/test_quant_contracts.py: concurrent ready branches, applicable join ordering, one activation per agent, typed UNAVAILABLE dependency completion, exception propagation, unrelated-root completion, transitive blocked state, and publication identity rejection. |
Passed; 29 focused tests | Controlled runners only. Real agent execution, typed input resolution/publication, run-record persistence, cost/state coordination, and UI consumption remain R02-R04/R06/R08/M01-M19. |
| A03a | b740cfe |
tests/test_scoped_rag.py, tests/test_scoped_file_writer.py, and tests/test_artifact_publisher.py: exact resolved current/baseline read paths are applied before ranking; metadata filters combine with the ACL; unauthorized paths, absent path metadata, and out-of-scope ticker chunks are rejected again after retrieval. Modified RAG modules also passed bytecode compilation. |
Passed; 21 tests, 1 symlink test skipped because directory symlink creation is unavailable on this Windows host | The bundled test runtime lacks the project's existing sentence_transformers dependency, so the legacy live embedding suite was not run. Migrated agent injection and the rest of A03 remain open. |
| R08a | 3d027e0 |
tests/test_quant_run_record.py plus routing, dependency, baseline, and full typed-contract regressions: terminal status is derived from agent records; frozen plan/session/run identity, baseline mode, input snapshot paths, and artifact inventory are validated; failed runs with no fabricated artifacts remain persistable; writes are atomic and immutable; malformed and legacy replay records remain untouched. |
Passed; 265 tests, 1 Windows symlink test skipped | Records are terminal-only and not yet emitted by the coordinator or consumed by the UI. Running lifecycle persistence and full R08 remain open. |
| R03a | 3753865 |
tests/test_published_inputs.py plus scoped writer, artifact publisher, I/O registry, and complete typed-contract regressions: publication identities are retained; required, optional, and baseline JSON are frozen separately; producer envelope schemas and reference/content identity are validated; unpublished and ACL-forbidden files are rejected; later optional publication does not mutate an earlier snapshot. |
Passed; 270 tests, 1 Windows symlink test skipped | The executor and concrete agent migrations do not yet invoke this resolver. Agent-specific ticker projection and input-model construction remain R03/M01-M19 rather than a generic JSON transform. |
| R03b | 487287b |
tests/test_dependency_executor.py plus routing, run-record, and complete typed-contract regressions: each ready batch captures completed dependency records before scheduling; sibling runners cannot mutate another activation's snapshot; legacy controlled runners remain compatible; capture failure is attributed to the agent and blocks only its dependents. |
Passed; 272 tests, 1 Windows symlink test skipped | The coordinator does not yet provide PublishedInputResolver or concrete input construction to this hook. Product runtime integration remains R03/R02/M01-M19. |
| M01a | 70f220c, e1c1f66 |
tests/test_market_data_normalizer.py, tests/test_datafetcher_typed.py, existing Finnhub tests, artifact publication/input resolution, and the full typed-contract regression set: the existing primary/fallback fetch returns without legacy writes for V8; actual OHLCV source is retained; invalid/missing values remain excluded, null, or explicitly unavailable; typed inputs, per-ticker JSON/Markdown, and the manifest publish through the ACL. |
Passed; 284 tests, 1 Windows symlink test skipped | Controlled providers only; no live credentials or calls. The legacy handler remains active and the coordinator does not yet invoke execute_typed, so M01 and product readiness remain open. |
| M02a | 50689bc |
tests/test_options_typed.py, options contracts, existing options-flow behavior, artifact publication/input resolution, and the expanded typed regression set: the existing bounded provider publishes sampled contracts, known-only totals, exact volume/OI/IV candidate criteria, unsupported provider capabilities, immutable JSON/Markdown, and explicit unavailable evidence for returned or raised provider failures. |
Passed; 227 tests, 1 Windows symlink test skipped | Controlled provider only; no live provider call. Candidate heuristics remain non-directional and unusual_flows stays empty because the current provider lacks trade-side, sweep/block, opening/closing, premium, and validated-classifier evidence. Baseline comparison and coordinator activation remain M02/runtime work. |
| M03a | e84d1e1, 2fd6704 |
tests/test_evidence_verifier.py, tests/test_trend_typed.py, the real structured-completion boundary with a controlled text response, analysis contracts, I/O bindings, providers, publication, scheduling, and the applicable typed regression set: current and baseline citations must resolve to frozen artifact identities and real JSON pointers/record IDs; the existing TrendAnalyst performs schema-valid LLM analysis, records usage when state is supplied, publishes immutable JSON/Markdown, compares prior trend context, and emits explicit unavailable evidence without LLM inference when OHLCV is absent. |
Passed; 264 tests, 1 Windows symlink test skipped | No live LLM request or coordinator activation. The repository-wide suite cannot collect because unchanged legacy academic files contain syntax errors and optional mcp, sentence_transformers, and bibtexparser dependencies are absent. Cost/trace coordination and product-path execution remain M03/R02/R03/R06 work. |
| M04a | cdb29c7 |
tests/test_sentiment_typed.py, the real structured-completion boundary with a controlled text response, sentiment and market-data contracts, I/O bindings, evidence verification, publication, scheduling, and the applicable typed regression set: the existing SentimentAnalyst preserves exact provider coverage and scored/unscored counts, rejects false neutral/sufficient claims for mostly unscored news, verifies current and baseline citations, compares approved prior sentiment, records usage when state is supplied, and publishes explicit unavailable analysis without LLM inference when news and earnings are absent. |
Passed; 268 tests, 1 Windows symlink test skipped | No live LLM request or coordinator activation. Trend context is optional and only consumed when its approved publication exists before activation. Cost/trace coordination and product-path execution remain M04/R02/R03/R06 work. |
| M05a | d2e84d0 |
tests/test_market_event_typed.py, the real structured-completion boundary with a controlled text response, market-event and market-data contracts, I/O bindings, evidence verification, publication, existing V6 options-flow behavior, and the applicable typed regression set: the existing MarketEventWatcher accepts only current DataFetcher-backed changes and exact copied news catalysts, rejects invented provider claims, requires cited approved prior context for delta claims, records usage when state is supplied, and publishes explicit unavailable context without LLM inference when all source records are absent. |
Passed; 273 tests, 1 Windows symlink test skipped | No live LLM request or coordinator activation. No separate external event provider is approved, so usable current context remains PARTIAL with EXTERNAL_EVENT_FEED_NOT_CONFIGURED; this is intentional rather than a request for a new integration. Product-path execution remains M05/R02/R03/R06 work. |
| M06a | cf7a68e |
tests/test_macro_context_typed.py, the real structured-completion boundary with a controlled text response, conditional market-event/options inputs, macro and options contracts, I/O bindings, evidence verification, publication, existing V6 options behavior, and the applicable typed regression set: the existing MacroContextWatcher preserves exact scope and upstream gaps, requires positive weighted drivers with current evidence, consumes approved prior context for cited delta comparison, records usage when state is supplied, and publishes UNCLEAR/UNKNOWN without LLM inference when no usable current evidence exists. |
Passed; 277 tests, 1 Windows symlink test skipped | No live LLM request or coordinator activation. Current MarketEventWatcher output is intentionally partial while no external event feed is approved, so that limitation propagates into macro availability. Product-path execution remains M06/R02/R03/R06 work. |
| M07a | cef6290 |
tests/test_unusual_event_typed.py, the real structured-completion boundary with a controlled text response, conditional market-event/macro/options inputs, event contracts, I/O bindings, evidence verification, publication, existing V6 options behavior, and the applicable typed regression set: the existing UnusualEventDetector separates meaningful, suppressed, and monitoring records; requires positive significance, known confidence, current evidence, stable global IDs, and cited prior context for carried delta events; records usage when state is supplied; and publishes monitoring-only unavailable output without LLM inference when current evidence is absent. |
Passed; 280 tests, 1 Windows symlink test skipped | No live LLM request or coordinator activation. A usable-evidence run with no meaningful event remains an LLM-owned classification, distinct from deterministic UNAVAILABLE. Product-path execution remains M07/R02/R03/R06 work. |
| M08a | 7e0ccc3 |
tests/test_event_router_typed.py, the real structured-completion boundary with a controlled text response, event-route and analysis contracts, exact current-event evidence verification, immutable publication, prompt constraints, and the applicable typed regression set: the existing EventRouter accounts for every meaningful event, permits explicit unassignment, rejects out-of-event-scope and zero-relevance assignments, records usage when state is supplied, and publishes an empty route without LLM inference when no meaningful event exists. |
Passed; 330 tests, 1 Windows symlink test skipped | No live LLM request or coordinator activation. EventRouter publishes evidence assignments only; the fixed routing plan remains the sole activation authority. Product-path execution remains M08/R02/R03/R06 work. |
| M09a-M12a | e44f80d |
tests/test_asset_analysts_typed.py, the real structured-completion boundary with controlled text responses, all four concrete asset schemas and registered expertise prompts, exact current and routed-event evidence verification, immutable per-agent publication, approved baseline loading/citation, upstream-gap preservation, deterministic unavailable output, and the applicable typed regression set. The legacy V6 specialized-node handoff test was also included. |
Passed; 339 tests, 1 Windows symlink test skipped | No live LLM request or coordinator activation. Shared execution plumbing preserves four distinct agents, fixed tickers, prompts, payload schemas, and artifacts; no shared asset-analysis file is written. Product-path execution remains M09-M12/R02/R03/R06 work. |
| M13a | 47e4355 |
tests/test_quantresearcher_typed.py, the real structured-completion boundary with controlled text responses, general and runtime-owned specialized conditional joins, current evidence verification, immutable publication, approved baseline loading, stable carried signal IDs and prior-record citations, upstream-gap preservation, deterministic unavailable output, and an LLM-owned empty signal result. The applicable typed regression set was included. |
Passed; 344 tests, 1 Windows symlink test skipped | No live LLM request or coordinator activation. Empty signals are accepted only after an actual evidence-bearing LLM pass; unavailable upstream analysis remains a distinct UNAVAILABLE artifact. Product-path execution remains M13/R02/R03/R06 work. |
| M14a | f504402 |
tests/test_backtester_typed.py, tests/test_decision_contracts.py, the real structured-completion boundary with a controlled text response, frozen signal/market-data resolution, exact signal-record citation, immutable publication, metric integrity, deterministic empty/unavailable behavior, and the applicable typed and YOLO/UI regression sets. |
Passed; 20 focused tests. Broader migrated typed suite and 44 YOLO/UI regressions passed; 1 Windows symlink test skipped. | No live LLM request, coordinator activation, or historical execution engine. Nonempty typed results are explicitly partial LLM_ASSESSMENT records, never simulated metrics. Product-path execution remains M14/R02/R03/R06 work. |
| M15a | c4fb7d7 |
tests/test_riskanalyst_typed.py, tests/test_decision_contracts.py, the real structured-completion boundary with a controlled text response, frozen backtest/context resolution, exact backtest-result citation, scope-filtered paper-position input, deterministic portfolio-heat validation, untested-signal rejection, immutable publication, completed all-rejected output, portfolio-only analysis, and deterministic empty/unavailable behavior. The applicable migrated typed regression set was included. |
Passed; 23 final focused tests; 308 broader tests passed with 1 Windows symlink test skipped before the final portfolio-only case was added. | No live LLM request or coordinator activation. The typed input uses normalized portfolio fractions; the later runtime adapter must convert existing percentage-point session state explicitly. Product-path execution remains M15/R02/R03/R06 work. |
| M16a | d2cca76 |
tests/test_quantstrategist_typed.py, tests/test_decision_contracts.py, the real structured-completion boundary with a controlled text response, frozen risk/context resolution, complete four-member cited voting, runtime-owned weight application, deterministic aggregation, exact risk validation, complete signal and proposal-lifecycle accounting, genuine all-rejected no-trade, proposal output, immutable publication, and deterministic unavailable strategy. The applicable migrated and existing committee regression set was included. |
Passed; 22 final focused tests. Expanded run collected 338 tests with 1 expected Windows symlink skip before the final lifecycle-negative case was added. | No live LLM request or coordinator activation. The agent records weights used but does not mutate or persist trusted performance state; that consistency belongs to I06. Product-path execution remains M16/R02/R03/R06 work. |
| M17a | d99c6fd |
tests/test_publisher_typed.py, tests/test_product_contracts.py, the real structured-completion boundary with controlled output, frozen strategy/context resolution, exact proposal rendering, genuine no-trade versus watch classification, unavailable-strategy labeling, pending-state validation, required current-strategy citation, immutable publication, and token accounting. The applicable migrated quant regression set was included. |
Passed; 15 focused tests; 333 broader tests passed with 1 expected Windows symlink skip. | No live LLM request, coordinator activation, or Slack delivery. The typed artifact contains the final Slack body, but the legacy product path remains active until Phase 5 integration. Product-path execution remains M17/R02/R03/R06/I01 work. |
| M18a | ce77b0b |
tests/test_vizagent_typed.py, scoped-writer, artifact-publisher, product-contract, and ACL parity tests: a declared matplotlib renderer produced a real PNG from typed OHLCV; LLM caption usage was recorded; image identity was bound to the visualization manifest and exact source artifact; unavailable executed-backtest metrics produced no chart; binary writes were atomic, ACL-scoped, and immutable. The applicable migrated quant regression set was included. |
Passed; 2 real-renderer tests; 47 focused support tests with 1 expected Windows symlink skip; 336 broader tests with 1 expected skip. | The configured test runtime required the project's already-declared matplotlib dependency to be installed temporarily. No coordinator/UI activation or live LLM request. VizAgent remains outside normal YOLO; product integration remains M18/I01/R03/R06 work. |
| M19a | e0c13b9 |
tests/test_portfoliomonitor_typed.py and tests/test_product_contracts.py: long/short percentage calculations, stop/target state, one fetch per ticker, exact position accounting, input immutability, quote provenance, partial and total quote loss, price-unavailable alerts, paused and closed-market runs, empty portfolios, provider exceptions, and rejection of altered position or threshold fields. The applicable migrated quant regression set was included. |
Passed; 17 focused tests; 343 broader tests passed with 1 expected Windows symlink skip. | Controlled quote provider only; no live provider call, separate monitor coordinator activation, state timestamp update, or product alert delivery. Those remain M19/I01/R06/R08 work. |
| C10, A02-A08, R01-R08, I01, I04-I06, U01-U07, V05, V06 | 30cc2cc |
The shared coordinator constructs one frozen plan, injects scoped ACL views, executes ready agents concurrently, resolves immutable inputs, publishes typed artifacts and evidence, persists replay/run state, exports the complete session, and exposes routing/permissions/status through the product API and UI. tests/test_quant_runtime.py and tests/test_v611_web_ui.py cover baseline/delta, same-day isolation, replay, export, unavailable/failed branches, and result-state distinctions. |
Passed locally on the active UI/coordinator path. | Live Slack and complete live Finley output remain V03/V04. |
| I02, I03, I07 | 748cf22 |
Existing quant commands and session behavior now dispatch the typed runtime by scope; quant product registration/help/README no longer present academic agents as active workflow nodes. | Passed in command, coordinator, scope-routing, and UI regressions. | Historical academic source files remain dormant and are excluded from the quant runtime; they are not V8 workflow nodes. |
| M18, M19, I01, U01 | d1a28b0 |
tests/test_quant_support_runtime.py and UI product-path tests invoke VizAgent and PortfolioMonitor through the shared coordinator, governed support runtime, ACL snapshots, evidence ledger, trace, and typed records. A real local PNG was served by the protected UI endpoint; controlled monitoring emitted an alert without mutating the position. |
Passed locally; 46 focused tests at implementation time. | M18 remains open for configured-LLM plus deployed-renderer verification; M19 remains open for live quote and delivery verification. |
| Runtime diagnostics | e5fcb2d |
Provider, executor, trace, UI, and agent failures redact query credentials, access tokens, and bearer values. A mocked Gemini 503 proves the request URL cannot enter persisted/user-facing errors. | Passed; 40 focused tests. | One historical pre-fix local failed-run record may contain an unsafe provider URL and was not deleted automatically. |
| Startup diagnostics | 388826c |
Slack startup reports configuration state without logging any signing-secret fragment or length; initialization exceptions pass through the shared credential sanitizer. | Passed; 27 app security and UI tests. | This verifies local diagnostics only; live Slack delivery remains V03. |
| Gemini authentication | 118def6 |
Gemini REST requests use the official x-goog-api-key header rather than a query parameter, so HTTP exception URLs cannot contain the provider credential. |
Passed; 60 LLM, structured-output, startup-security, and executor tests. | Provider quota remains external; this change does not satisfy live-agent gates. |
| M01 | 5f4ae68 |
Missing Finnhub credentials now skip Finnhub candles and retain the existing yfinance OHLCV fallback; news/earnings remain explicit unavailable components. Live UI runs 56624454, c5684180, 72c8a536, and 3db04eef published partial DataFetcher artifacts with usable yfinance OHLCV and completed OptionsFlowWatcher output. |
Passed focused provider tests and configured live UI execution. | Local .env has no Finnhub key, so live news remained unavailable by design. |
| C09, M03, M05 | 589695d |
Gemini structured calls use native JSON mode while retaining strict Pydantic validation and one correction; prompts state baseline/delta and exact evidence rules. Live run c5684180/20260927T044130245818Z-9a35604c published cited TrendAnalyst output with recorded usage. Live run 72c8a536/20260927T044501770263Z-2a9d66db published cited TrendAnalyst and MarketEventWatcher outputs with recorded usage. Retry c3e037fb/20260927T045929916277Z-0a1982bd again published MarketEventWatcher while recording the provider failures and blocked dependents by name. |
Passed focused tests and configured live LLM execution through both agents. | The retry received a Gemini rate-limit response for TrendAnalyst and HTTP 503 for MacroContextWatcher; downstream agents could not run, so M06-M17 remain open. |
| U08, V01, V02 | 8.0 through 118def6 |
Applicable quant/provider/coordinator/RAG/session/UI regression: 408 passed, 1 Windows symlink test skipped, 1 documented stale legacy strategist test deselected. Built polyhedra-8.0.0-py3-none-any.whl; verified all 103 packaged files include the governed contracts, primary/support runtimes, sanitizer, and UI assets. Local startup initialized the existing Slack handler. Browser checks at 1440 x 900 and 390 x 844 exercised Run, Routing Table, and Permissions without overflow. |
Passed locally. | Repository-wide collection still includes retired academic modules outside the quant product. No HF deployment was requested for this audit. |
| M06, prompt boundary | 7ba97b7, 4945749 |
Live UI session 6bb08b7f, run 20260929T044302718680Z-be7902ba, completed DataFetcher, OptionsFlowWatcher, TrendAnalyst, MarketEventWatcher, and MacroContextWatcher through the governed runtime. MacroContextWatcher published weighted cited drivers with trace usage (35,061 prompt tokens, 4,175 completion tokens, 18.7 seconds). The prompt projector retains the complete validated input snapshot and full options artifact while omitting raw contract rows and repeated nested candidate evidence identities only from MacroContextWatcher/UnusualEventDetector prompts; focused tests verify type revalidation, non-mutation, registry scope, and retained aggregates/candidates. The final projection reduced measured MacroContextWatcher and UnusualEventDetector prompt sizes from 105,213/114,210 bytes to 57,022/66,019 bytes. |
Passed; 23 focused tests and 411 applicable regression tests passed, with 1 expected Windows symlink skip and 1 documented stale legacy strategist test deselected. | UnusualEventDetector reached the provider after MacroContextWatcher but received HTTP 503. Two later live attempts received HTTP 503 on the first analyst batch, so M07-M17 and full Finley output remain open. Full typed input snapshots and exported evidence are not compacted. |
| BTC route and direct evidence | 461eb1f |
Live UI session f3ad4c3c, run 20260929T045754944935Z-78667725, produced the exact BTC specialized plan without OptionsFlowWatcher. DataFetcher, MarketEventWatcher, and MacroContextWatcher completed while a sibling failure did not stop the independent event branch. Trace evidence showed TrendAnalyst's completed LLM output was rejected for an out-of-input artifact identity, while UnusualEventDetector received HTTP 503. TrendAnalyst now explicitly cites exact per-ticker MarketDataArtifact identities for value paths, and UnusualEventDetector explicitly cites only its direct current watcher dependencies; focused tests assert both rules. A fresh corrected run 2e595e3b/20260929T050357093085Z-1590b037 preserved BTC options exclusion but Gemini returned rate-limit errors on the first analyst batch. |
Passed; 15 focused tests and 411 applicable regression tests passed, with 1 expected Windows symlink skip and 1 documented stale legacy strategist test deselected. | BTC routing/options exclusion is verified, but M09 and V04 remain open because BTCAnalyst and Finley did not execute successfully. No provider retry policy was added because it is outside the approved PRD and awaits user approval. |
| Final local audit | a03fcd1 |
Re-ran the complete applicable quant/provider/coordinator/session/UI suite after the shared direct-evidence prompt rule, then rebuilt the distributable wheel. Audited every unchecked function point against the verification record. | 411 passed, 1 expected Windows symlink skip, 1 documented stale legacy strategist test deselected; polyhedra-8.0.0-py3-none-any.whl built successfully. |
Seventeen parent gates remain deliberately unchecked: M04, M07-M19, V03, V04, and V07. They require stable configured live LLM/provider execution, live Slack/UI support verification, and a requested HF deployment. Two bounded retries for transient LLM 429/503/timeouts were proposed but not implemented because they require product approval under the project guardrails. |
| Session storage configuration | 418081b |
Verified that the documented POLYHEDRA_SESSIONS_DIR setting controls the default SessionManager storage root used by Slack/UI while an explicit constructor path retains precedence. Re-ran the focused session/UI suite and complete applicable V8 regression suite. |
48 focused tests passed; 412 applicable tests passed, with 1 expected Windows symlink skip and 1 documented stale legacy strategist test deselected. | No live laptop run was used as release evidence. Live Slack/UI/provider and requested HF deployment gates remain open. |