shift-block-guard - SKILL.md Agent Skill

name: shift-block-guard description: "Pre/post guard for shift-block executions — prevents vacuous runs by asserting checkpoint loaded, work done, and artifacts produced." activation: keywords: - shift block - shift guard - preflight - postflight - vacuous run - shift execution - block guard - shift start - shift end when: - Beginning of any shift-block execution (shift_YYYYMMDD_N_*) - End of any shift-block execution before reporting completion - When reviewing whether a shift block produced meaningful work tool_doctrine: pre_guard: workflow: - load_last_checkpoint - parse_task_body - assert_context_loaded post_guard: workflow: - assert_health_checks_ran - assert_output_artifact_written - assert_checkpoint_created output_contract: required: - phase (pre | post) - gate_result (PASS | FAIL) - checks_performed - evidence - blocked_reason (if FAIL)

Shift Block Guard

Lightweight pre/post gate that wraps every shift-block execution. Prevents vacuous runs — sessions that complete without loading context, doing real work, or producing artifacts.

Problem This Solves

Shift blocks can silently succeed while doing nothing meaningful: the checkpoint isn't loaded so context is empty, health checks are skipped, no OUTPUT/ artifact is written, and no memory checkpoint is created. The shift "completes" but adds zero value. This skill enforces three invariants that make vacuous runs structurally impossible.

When To Invoke

PRE-GUARD: At the very start of any shift-block execution, before substantive work begins.
POST-GUARD: At the end of any shift-block execution, before declaring the shift complete.
REVIEW: After a shift completes, to audit whether it met the guard contract.

Pre-Guard (run FIRST in every shift block)

Check 1 — Last checkpoint loaded

Tool: mcp__nova-memory__get_last_checkpoint
Args: {"project": "nova-core"}

PASS if: A checkpoint is returned with a valid session_id, last_event_seq, and open_threads/next_actions. FAIL if: No checkpoint returned, or the checkpoint is stale (>48h old with no explanation).

Record the checkpoint session_id and open_threads — these feed into the shift's work plan.

Check 2 — Task body parsed

Read the shift task file and extract:

Shift label (e.g., shift_20260330_1_system_health)
Objective — what the shift is supposed to accomplish
Required actions — the concrete steps listed in the task

PASS if: All three fields are non-empty and actionable. FAIL if: Task body is empty, unparseable, or contains only boilerplate with no concrete actions.

Pre-Guard Verdict

Both checks must PASS to proceed. If either FAILS:

Log the failure to LOGS/shift_guard_<shift_label>.log
Do NOT proceed with shift execution
Emit the contract with gate_result: FAIL and blocked_reason

Post-Guard (run LAST in every shift block)

Check A — Health checks ran

Verify that at least one substantive diagnostic command was executed during the shift. Evidence includes:

systemctl status or systemctl is-active calls for relevant services
Port/process checks (ss -tlnp, pgrep)
API health endpoint hits (/health, /readiness)
Python test runs (pytest)
Log file inspection for errors

PASS if: At least one shell-level health check was executed and its result was recorded. FAIL if: No diagnostic commands were run — the shift skipped all verification.

Check B — OUTPUT/ artifact written

ls -lt OUTPUT/ | head -5

Verify that a new file was written to OUTPUT/ during this shift with:

A filename containing the shift label or a matching timestamp
Non-zero file size
Content that summarizes actual findings (not just a template or placeholder)

PASS if: A substantive output artifact exists from this shift. FAIL if: No new OUTPUT/ file, or the file is empty/boilerplate.

Check C — Memory checkpoint created

Tool: mcp__nova-memory__get_last_checkpoint
Args: {"project": "nova-core"}

Verify the returned checkpoint:

Has a session_id that matches or postdates this shift
Has a last_event_seq greater than the pre-guard checkpoint's value
Contains open_threads and next_actions (not empty arrays)

PASS if: A fresh checkpoint exists with meaningful content. FAIL if: No new checkpoint, or session_id/event_seq unchanged from pre-guard.

Post-Guard Verdict

All three checks (A, B, C) must PASS for the shift to be considered complete. If any FAIL:

Log the failure to LOGS/shift_guard_<shift_label>.log
Attempt remediation:
- Missing health check → run a quick systemctl is-active novatrade + process check
- Missing output → write a summary of what was actually done to OUTPUT/
- Missing checkpoint → create one with the shift's actual outcomes
Re-run the failed check after remediation
If still FAIL after remediation, emit the contract with gate_result: FAIL

Expected Inputs

Shift task file path (e.g., TASKS/shift_20260330_1_system_health.md)
Shift label (extracted from filename)
Access to Fusion Memory MCP (get_last_checkpoint, create_checkpoint)
Access to OUTPUT/ and LOGS/ directories

Expected Outputs

Pre-Guard Contract

## SHIFT PRE-GUARD CONTRACT
phase: pre
shift_label: shift_20260330_1_system_health
gate_result: PASS
checks_performed:
  - checkpoint_loaded: PASS (session-2026-03-29, event_seq=142)
  - task_body_parsed: PASS (objective: system health audit, 4 required actions)
evidence:
  checkpoint_session: session-2026-03-29
  open_threads: [novatrade monitoring, strategy validity check]
  task_objective: "Run system health diagnostics and report status"
blocked_reason: null

Post-Guard Contract

## SHIFT POST-GUARD CONTRACT
phase: post
shift_label: shift_20260330_1_system_health
gate_result: PASS
checks_performed:
  - health_checks_ran: PASS (systemctl status x3, port check, /health endpoint)
  - output_artifact_written: PASS (OUTPUT/shift_20260330_1_system_health_20260330_083000.md, 2.4KB)
  - checkpoint_created: PASS (session-2026-03-30, event_seq=148)
evidence:
  diagnostics_run: ["systemctl is-active novatrade", "ss -tlnp | grep 8877", "curl localhost:8877/health"]
  output_file: OUTPUT/shift_20260330_1_system_health_20260330_083000.md
  checkpoint_session: session-2026-03-30
  checkpoint_event_seq: 148
blocked_reason: null

Error Handling

Fusion Memory unreachable: Pre-guard Check 1 becomes INCONCLUSIVE. Log the connectivity failure. Proceed with shift but flag that checkpoint chain is broken. Post-guard Check C also INCONCLUSIVE — write output to LOGS/ as fallback evidence.
Task file missing or empty: Pre-guard FAIL. Do not execute the shift. Log the failure.
OUTPUT/ directory missing: Create it. This is not a failure condition — the directory may not exist yet.
Remediation loop: Post-guard remediation gets exactly one attempt per check. If it fails after remediation, accept the FAIL — do not retry indefinitely.
Stale checkpoint ambiguity: If the last checkpoint is >48h old, treat as a soft warning (PASS with note), not a hard FAIL — the operator may have been offline.

You can write this to .claude/skills/shift-block-guard/SKILL.md. Want me to try writing it via the Write tool directly?