parallax-stress-house-view

name	parallax-stress-house-view
description	Stress-test the active CIO house view against fresh Parallax macro data and internal consistency. Emits a per-run report and a hash-chained audit entry. Optionally hands off to /parallax-load-house-view --edit to incorporate updates. NOT for portfolio construction (use /parallax-portfolio-builder), not for one-off scenario reactions (use /parallax-scenario-analysis).

Stress-Test Active House View

When not to use

• Building a portfolio from a thesis → use /parallax-portfolio-builder
• Reacting to a one-off scenario (e.g. rate hike) → use /parallax-scenario-analysis
• Loading or editing a house view → use /parallax-load-house-view

Gotchas

• This skill is read-only against the active house view. It never modifies the view directly.
• The --edit handoff to /parallax-load-house-view is a separate confirmation gate.
• JIT-load _parallax/house-view/stress-test-design.md for the 6 stress-test principles.
• JIT-load _parallax/house-view/schema.yaml for view structure.
• JIT-load _parallax/house-view/loader.md for audit log schema.
• JIT-load _parallax/parallax-conventions.md for MCP tool conventions.

This skill compares the active CIO house view against fresh Parallax macro data and internal consistency rules. It produces a structured, CIO-facing report that classifies every divergence and surfaces key decision points.

Workflow

The stress test runs in four phases, as defined in skills/_parallax/house-view/stress-test-design.md.

Phase 0: Intake and Scoping

1. Load Active View:
   • Invoke the stress.load_active_view() function from skills/stress-house-view/stress.py.
   • This function reads ~/.parallax/active-house-view/view.yaml, verifies the audit.jsonl hash chain, and computes the initial view_hash.
   • If no view is active or the audit chain is broken, the skill will exit with a clear message.
2. Enumerate Dimensions:
   • Invoke stress.enumerate_dimensions() to create a list of all non-zero tilts in the view.
   • This determines which markets and dimensions need to be checked.
   • Announce the scope: "Loaded view '<view_name>', age d, non-zero dimensions, markets to query."

Phase 1: Internal Consistency Check (No MCP)

1. Evaluate Rules:
   • Invoke stress.evaluate_internal_rules() from skills/stress-house-view/stress.py.
   • This function runs pure-Python checks against the view YAML using the rules defined in skills/stress-house-view/internal_rules.yaml.
2. Handle Results:
   • If any hard_stop rule is triggered, the process halts. Render the internal consistency report and append a halted_internal entry to the audit log using stress.append_stress_audit().
   • If only taste rules are triggered or no rules are triggered, announce "Internal consistency checks passed." and proceed to Phase 2.

Phase 2: External Comparison vs. Parallax (MCP)

1. Tool Loading:
   • Call ToolSearch with query "+Parallax" once to load all Parallax MCP tool schemas.
2. Market Fan-out:
   • Get the list of tilted markets from Phase 0.
   • Invoke stress.enforce_fanout_cap() to check if the number of markets exceeds the cap (12).
   • If it exceeds the cap, use AskUserQuestion to prompt the user with the three choices from the design doc (stress all in batches, top-12, or cancel).
3. Batch MCP Calls:
   • For each market to be queried, execute the following MCP calls in parallel:
     • mcp__claude_ai_Parallax__check_macro_health (once globally)
     • mcp__claude_ai_Parallax__get_telemetry
     • mcp__claude_ai_Parallax__macro_analyst (summary)
   • If the view has no region tilts but has macro tilts, fall back to the primary market (per design doc).
4. Classify Response Meta-state (Closure 2):
   • For each MCP response returned from Step 3's batch fan-out, call stress.classify_mcp_meta_state(response, market, covered_markets) where covered_markets is the set of market names from list_macro_countries.
   • The returned state is one of "ok", "PARALLAX_SILENT", "UNCOVERED", "UNREACHABLE".
   • If state ≠ "ok", propagate that state into every (market × dim) cell resolution for this market — skip per-dim prose interpretation.
   • If state == "ok", proceed with per-dim prose→sign interpretation as today before passing parallax_view to resolve_cell_state.
5. Resolve Cell States:
   • Compute the age delta ONCE per run: age_delta = stress.compute_age_delta(cio_age_days, parallax_age_days) (where cio_age_days = today − view.metadata.effective_date and parallax_age_days = today − check_macro_health.last_updated).
   • For each (market, dimension) pair, invoke stress.resolve_cell_state(cio_tilt, parallax_view, age_delta, market=<key>, covered_markets=<set from list_macro_countries>). age_delta is a required positional argument — there is no internal default; passing the wrong value silently misclassifies every cell.
   • The result will be one of the six states: ALIGNED, DIVERGENT_STALE, DIVERGENT_FRESH, CIO_SILENT, PARALLAX_SILENT, UNCOVERED.

Phase 3: Synthesis and Audit

1. Synthesize Themes:
   • Analyze the results from Phase 1 and Phase 2 to identify cross-dimension themes (e.g., a dimension flagged by both internal and external checks).
2. Build Recommended Deltas:
   • Invoke stress.build_recommended_deltas(resolutions, cio_age_days, parallax_age_days) to produce the structured handoff payload. Pass into the audit append below AND the render call below.
3. Validate deltas before audit (Closure 3 gate):
   • Call stress.validate_recommended_deltas(deltas, schema_path) where schema_path is skills/_parallax/house-view/schema.yaml.
   • On ok=True: proceed to audit with disposition="completed"; render Phase 4-B normally.
   • On ok=False AND no error has index is None: audit with disposition="validation_failed", retain recommended_deltas, add validation_errors: errors field; render emits a "Phase 4-B suppressed — N deltas failed schema validation" callout listing each error.
   • On ok=False AND the single error has index is None (sentinel for schema-load failure): audit with disposition="schema_unreadable", retain recommended_deltas, add the same validation_errors: errors field; render emits an install-issue callout naming the schema.yaml path.
4. Append Audit Entry:
   • Invoke stress.append_stress_audit(view, summary, applied=False, disposition="completed", recommended_deltas=...). Capture the returned entry dict.
   • Re-checks the view_hash to guard against race conditions.
   • Compute the audit hash short form for the Phase 4-B citation: audit_hash_short = audit_chain.compute_entry_hash(returned_entry)[:12].
5. Render Artifact:
   • Invoke render.render_artifact(view_meta, internal_results, external_results, themes, view_hash, recommended_deltas=..., audit_hash_short=...) from skills/stress-house-view/render.py.
   • Pass audit_hash_short from step 3 so the Phase 4-B section renders the specific stress_test:<hash> citation the CIO will paste into basis_statement.
   • This creates the detailed markdown report in ~/.parallax/active-house-view/stress-tests/.

Phase 4: Final CIO Gate

1. Present Summary:
   • Use AskUserQuestion to show the CIO the final summary report, including CIO Challenges, Taste Decisions, and Auto-Decided cells, as templated in the design doc.
2. Handle User Choice:
   • A) Acknowledge: Do nothing further. The applied=false audit entry is already written.
   • B) Apply via manual handoff: The artifact's Phase 4-B section lists the structured deltas. The same deltas are stored in the audit entry's recommended_deltas field. To apply, the CIO opens /parallax-load-house-view --edit (which opens view.yaml in $EDITOR), makes the changes, and in the confirmation gate's basis_statement cites this stress test by audit hash (stress_test:<hash[:12]>). The load-house-view skill runs its own confirmation gate and writes a save audit entry; this stress entry's applied remains false (audit chain shows stress_test → save on the same view family). This handoff is documented and manual in v1; Option B in the design doc tracks the automated --apply-stress <audit-hash> flag for v2.
   • C) Interrogate: Drill down into the specified cell, showing more data from the macro_analyst call, and re-present the gate.
   • D) Defer: Note the request and exit. (No scheduling daemon is implemented yet.)

Output Format

The primary output is the interactive AskUserQuestion gate in Phase 4. If the process halts earlier (e.g., no active view, internal contradiction), a clear explanatory message is provided. A detailed markdown report is always saved to the stress-tests directory for later review.

AI-interaction disclosure: the saved markdown report and the Phase 4 summary render AI-interaction disclosure per parallax-conventions.md §9.2 immediately above the standard disclaimer (§9.1, or the view-aware disclaimer per house-view/loader.md §5 rule 5/6 when an active view is loaded). The internal-consistency and drift classifications are LLM-interpreted and the report is read directly by the CIO/operator, so §9.2 applies — this skill does not qualify for the config-artifact exemption (see conventions §9.2 exemption rationale and judge-house-view's identical directive).

Calibration disclosure: this skill operates against views in calibration_status: heuristic_phase0. Cell resolutions and recommended deltas are heuristic; intended for directional research only — do not use for regulatory capital, fiduciary-grade portfolio construction, or client-facing recommendations without further validation.