Live Continuous 3-Run Proof

This page is the public evidence for the reopened roadmap claim: AgentXchain completed three back-to-back governed runs from a project-local VISION.md with no manual accept-turn, no manual gate approval, no HUMAN-ROADMAP edits mid-session, and at least one real-credential turn per run.

Date of proof: 2026-04-17

What It Is

• Category: live proof example
• Surface: agentxchain run --continuous
• Runtime shape: local_cli authoring roles + real api_proxy QA review
• Credentialed provider: Anthropic (claude-haiku-4-5-20251001)
• Source harness: examples/live-governed-proof/run-continuous-3run-proof.mjs

Workflow Or Proof Shape

The proof used a temp governed repo with a project-local VISION.md containing three distinct goals:

• capture alpha ledger evidence trail
• capture beta provenance dashboard slice
• capture gamma escalation recovery drill

Each governed run used the truthful mixed-runtime shape:

• pm and dev authored repo-local artifacts through local_cli
• qa ran through a real Anthropic-backed api_proxy runtime in review_only mode
• run --continuous derived the next intent from VISION.md, auto-triaged it via --triage-approval auto, and advanced through planning, implementation, and QA without operator input

run-agents.sh is now documented as the raw fallback path. run --continuous is the primary product surface for lights-out operation.

Exact Command

agentxchain run --continuous --vision .planning/VISION.md --max-runs 3 --poll-seconds 30 --triage-approval auto

Observed harness invocation:

set -a; source .env; set +a
node examples/live-governed-proof/run-continuous-3run-proof.mjs --json --keep-temp

Proof workspace:

/var/folders/2c/vqqd_9tn4wd_flr5f0690yl80000gn/T/axc-live-cont-3run-BbcBIi

Evidence Summary

Metric	Value
Session ID	cont-0e280ba0
Started	2026-04-17T16:48:55.218Z
Finished	2026-04-17T16:49:49.796Z
Wall clock	54.578s
Runs completed	3
Final session status	completed
Total cumulative spend	$0.025
Provider / model	anthropic / claude-haiku-4-5-20251001
Blockers hit	none
VISION.md changed	no

Session Snapshots

Before the command started, .agentxchain/continuous-session.json did not exist.

Initial snapshot after the session materialized:

{
  "session_id": "cont-0e280ba0",
  "started_at": "2026-04-17T16:48:55.518Z",
  "vision_path": ".planning/VISION.md",
  "runs_completed": 0,
  "max_runs": 3,
  "idle_cycles": 0,
  "max_idle_cycles": 3,
  "current_run_id": null,
  "current_vision_objective": null,
  "status": "running",
  "per_session_max_usd": null,
  "cumulative_spent_usd": 0,
  "budget_exhausted": false
}

Final snapshot:

{
  "session_id": "cont-0e280ba0",
  "started_at": "2026-04-17T16:48:55.518Z",
  "vision_path": ".planning/VISION.md",
  "runs_completed": 3,
  "max_runs": 3,
  "idle_cycles": 0,
  "max_idle_cycles": 3,
  "current_run_id": "run_920d1193b9521850",
  "current_vision_objective": "Proof Queue: capture gamma escalation recovery drill",
  "status": "completed",
  "per_session_max_usd": null,
  "cumulative_spent_usd": 0.025,
  "budget_exhausted": false
}

Per-Run Evidence

Turn counts below are derived from accepted .agentxchain/history.jsonl entries grouped by run_id, which is the truthful per-run count for this session.

Run	Vision goal	Trigger	Model used	Accepted turns	Spend	QA artifact
run_2d751d643cf6ba01	Proof Queue: capture alpha ledger evidence trail	vision_scan	claude-haiku-4-5-20251001	3	$0.008	.agentxchain/reviews/turn_4d196165f84b4485-qa-review.md
run_2f48ef8a12614d13	Proof Queue: capture beta provenance dashboard slice	vision_scan	claude-haiku-4-5-20251001	3	$0.008	.agentxchain/reviews/turn_dcc77f924f63955b-qa-review.md
run_920d1193b9521850	Proof Queue: capture gamma escalation recovery drill	vision_scan	claude-haiku-4-5-20251001	3	$0.009	.agentxchain/reviews/turn_8f0fcbf4de0f9d6f-qa-review.md

Every run carried the same provenance shape:

• trigger: "vision_scan"
• created_by: "continuous_loop"
• a distinct intake_intent_id

Git Log Evidence

git log --oneline from the proof repo shows run-derived commits for all three runs, and every run-derived commit includes the session trailer Continuous-Session: cont-0e280ba0.

54689bc proof(implementation): run_920d1193b9521850 proof-queue-capture-gamma-escalation-recovery-drill
f8279ea proof(planning): run_920d1193b9521850 proof-queue-capture-gamma-escalation-recovery-drill
84508d8 proof(implementation): run_2f48ef8a12614d13 proof-queue-capture-beta-provenance-dashboard-slice
66b5cf5 proof(planning): run_2f48ef8a12614d13 proof-queue-capture-beta-provenance-dashboard-slice
2d8506a proof(implementation): run_2d751d643cf6ba01 proof-queue-capture-alpha-ledger-evidence-trail
5f17e1e proof(planning): run_2d751d643cf6ba01 proof-queue-capture-alpha-ledger-evidence-trail
df74d41 init proof repo

Example commit trailer:

Continuous-Session: cont-0e280ba0
Co-Authored-By: GPT 5.4 (Codex) <[email protected]>

Notes

• No blocker or escalation fired during the clean run.
• No manual gate approval was needed because the continuous loop executed with auto-approval enabled for gate handling.
• The project-local VISION.md remained unchanged throughout the session.
• A pre-proof harness hardening pass was required earlier on 2026-04-17; that exploratory run is not counted as the public proof. The evidence on this page is from the subsequent clean rerun above.

Run It

export ANTHROPIC_API_KEY=...
node examples/live-governed-proof/run-continuous-3run-proof.mjs --json --keep-temp

The harness creates a temp governed repo, runs the exact run --continuous command shown above, and prints the evidence as JSON. Use --keep-temp when you want to inspect the proof repo's .agentxchain/ artifacts and git history afterward.

Key Takeaways

• The strongest marketing claim now has public evidence: one unattended continuous session completed three governed runs with real provider-backed QA review.
• run --continuous is the primary lights-out path; run-agents.sh is only the raw fallback.
• The proof leaves durable provenance in three places at once: continuous-session.json, run-history.jsonl / history.jsonl, and git commit trailers.