The agent's decisions
When a submission fails, the agent gathers evidence with read-only tools, tests a hypothesis by simulation, and commits one cause-appropriate remedy, bounded by a deterministic guardrail.
Steering the live retry above. The agent learns from each failed tip and escalates until the bundle lands.
Attempt 1Raise tip
conf 0.82Fee starvation — tip below the effective (rising) tip floor caused the bundle to be dropped without landing
landed=false and classified_kind=Dropped indicate a pure inclusion failure, not a validity failure. Blockhash age is 35 blocks, well under the ~150 expiry, so refresh is unnecessary. CU consumed in simulation is 450 vs 10000 requested, ruling out compute exhaustion. Simulation returns ok with err=null, ruling out adverse-market/program error. The tip paid (1823) equals the static p50 exactly, but the EMA p50 has drifted up to 3794 and p75 is 4045 — the floor is rising and p50 is no longer competitive. Sim at 4045 also succeeds. Therefore raise the tip toward p75/EMA.
blockhash_age_blocks 35classified_kind Droppedcu_requested 10,000error not confirmed within timeoutlanded falsetip_floor_p50 1,823tip_paid_lamports 1,823
evidence gathered: get_failure_context, get_network_conditions, get_tip_floor, get_retry_history, simulate_with_params ×2
show reasoning trace
I'll diagnose this systematically. Let me gather evidence from multiple sources.
Key observations:
- **Blockhash age: 35 blocks** — blockhashes expire at ~150 blocks, so this is NOT expired. Refresh alone won't help.
- **landed: false, classified "Dropped"** — never landed.
- **tip_paid = 1823 = exactly p50**, but the **EMA p50 is 3794** (more than double the static p50), and p75 = 4045. This signals the tip floor is rising; paying exactly p50 is borderline/insufficient — classic fee starvation.
- **CU not exhausted** (consumed isn't even reported as near-limit; this is a drop, not a compute failure).
This looks like **fee starvation**: tip at p50 (1823) while EMA has drifted to 3794. Let me test the hypothesis before committing.
The simulation confirms the transaction itself is valid (no program/market error, CU consumed only 450 vs 10000 requested — definitely not compute exhaustion). The drop is purely a landing/inclusion problem, consistent with fee starvation.
Let me confirm the diagnosis pieces:
- **Not blockhash expiry**: age 35 << 150 expiry threshold.
- **Not compute exhaustion**: CU consumed 450 << 10000 requested.
- **Not adverse-market/program error**: sim returns ok, err null.
- **Fee starvation**: tip paid exactly p50 (1823) while EMA p50 has drifted up to 3794 and p75 is 4045 — the floor is rising and p50 is no longer competitive enough to land.
Remedy: raise the tip toward a higher percentile (p75 = 4045, which is also near the EMA of 3794).
Attempt 2Raise tip
conf 0.82Fee starvation — bundle repeatedly dropped because tip is below the effective inclusion floor
Two prior attempts never landed: attempt 1 paid 1823 (=p50) and attempt 2 paid 4045 (=p75), both classified Dropped/not_landed. Blockhash age is only 33 blocks (well within validity), so no refresh needed. Simulation shows cu_consumed=450 against cu_requested=10000, so compute exhaustion is impossible. Simulation returned ok=true/err=null, so this is not an adverse-market or program error — the tx is valid and would execute if included. The tip-floor EMA (3794) sits above p50, confirming a rising floor, and p75 has already failed to land. Escalating toward p95 (96837) is the appropriate remedy for inclusion failure.
blockhash_age_blocks 33classified_kind Droppedcu_requested 10,000error not confirmed within timeoutlanded falsetip_floor_p50 1,823tip_paid_lamports 4,045
evidence gathered: get_failure_context, get_network_conditions, get_tip_floor, get_retry_history, simulate_with_params ×2
show reasoning trace
I'll diagnose this systematically. Let me gather evidence from multiple sources.
Let me analyze the evidence:
- **Blockhash age: 33 blocks** — still well within the ~150 block validity window, so NOT a blockhash expiry issue.
- **CU requested 10000, landed=false** — never landed, so we can't compare CU consumed vs requested for exhaustion. But 10000 CU is tiny; no exhaustion error signal.
- **Tip history**: attempt 1 paid 1823 (=p50) → dropped. Attempt 2 paid 4045 (=p75) → dropped. Both never landed.
- **Tip floor**: p50=1823, p75=4045, p95=96837, EMA=3794. Current tip (4045) is at p75 but below the EMA-adjusted competitive level and far below p95.
This looks like **fee starvation**: the bundle never lands, and escalating from p50→p75 didn't help. The EMA (3794) sitting above p50 suggests the floor is rising. I need to test a higher tip toward a higher percentile.
Let me simulate hypotheses before committing.
Important findings from simulation:
- **CU consumed: 450** vs requested 10000 — massive headroom. This is definitively NOT compute exhaustion. CU is fine.
- Simulation succeeds (ok=true, err=null) — so this is NOT an adverse-market/program error. The transaction is valid and would execute cleanly if it lands.
The problem is purely a **landing/inclusion** problem. The tx is valid (sim ok), CU is generous, blockhash is fresh (33 blocks). It simply never gets included despite two attempts. This is classic **fee starvation** — the tip needs to escalate toward a higher percentile.
Attempt 1 (p50=1823) and attempt 2 (p75=4045) both failed to land. The EMA (3794) confirms the effective floor is above p50. The next meaningful escalation step is toward p95 (96837), since p75 already failed. I'll commit to raising the tip.
Controlled fault injection: three different failures, three different remedies.
Blockhash expiry (TransactionExpiredBlockheightExceeded)
blockhash_age_blocks=162 exceeds the ~150-block validity window, and cu_consumed=0 proves the tx never executed (not compute exhaustion). The error is the canonical TransactionExpiredBlockheightExceeded. Tip is not the problem: 50000 lamports already exceeds tip-floor p50=30000 and EMA=22696 under 'normal' congestion, so no tip raise is warranted. Simulation with refresh_blockhash=true succeeds (ok=true) at the unchanged tip; raising cu_limit to 220000 left consumption fixed at 199950, confirming a real ~50-CU margin rather than a ceiling collision. The remedy is purely a blockhash refresh with no tip change.
blockhash_age_blocks 162cu_consumed 0cu_requested 200,000error TransactionExpiredBlockheightExceededlanded falsetip_paid_lamports 50,000
evidence gathered: get_failure_context, get_network_conditions, get_retry_history, get_tip_floor, simulate_with_params ×3
show reasoning trace
I'll diagnose this failure by gathering evidence from the available tools.
The evidence points clearly to blockhash expiry:
- `blockhash_age_blocks: 162` — far beyond the ~150-block validity window, so the blockhash is definitively expired.
- `cu_consumed: 0` — the transaction never executed, ruling out compute exhaustion (would need CU consumed ≈ requested).
- `error: TransactionExpiredBlockheightExceeded` — the canonical expiry signal.
- `tip_paid_lamports: 50000` — already above p50 (30000) and well above the EMA (22696). Congestion is "normal." This is NOT fee starvation; the tip is healthy.
Let me test the refresh hypothesis before committing.
Wait — the simulation shows `cu_consumed: 199950` against a `cu_limit` of 200000. That's a razor-thin 50-CU margin (99.975% utilization). While the refresh succeeds in simulation, that margin is dangerously tight; under real execution the transaction could tip into compute exhaustion. Let me verify whether the tight margin is real and test a small CU bump.
Both simulations succeed and consume a stable 199950 CU regardless of the limit. The consumption is fixed at 199950 — it's not bumping against the ceiling (raising the limit to 220000 didn't change consumption), so this is genuine real CU usage with a real, if thin, 50-CU headroom under the original 200000 limit. The transaction executes fine; the only thing that killed it was the expired blockhash.
The root cause is unambiguously **blockhash expiry**. The correct remedy is to refresh the blockhash and leave the tip unchanged.
Fee starvation — tip below the Jito tip floor caused the bundle to never be included
tip_paid_lamports=8000 is below even tip floor p25=12300 and far below p50=30000; landed=false with inflight_status=Pending and cu_consumed=0 means the tx never executed (not compute exhaustion) and error=null rules out a program/market abort. blockhash_age_blocks=40 is within the ~150-block validity window, so no refresh is needed. Simulation at 91863 (p75) succeeded (ok=true), as did p50; I chose p75 for reliable inclusion margin above the median floor without overpaying toward p95=549095.
blockhash_age_blocks 40cu_consumed 0cu_requested 200,000error nullinflight_status Pendinglanded falsetip_floor_p50 30,000tip_paid_lamports 8,000
evidence gathered: get_failure_context, get_network_conditions, get_tip_floor, get_retry_history, simulate_with_params ×2
show reasoning trace
I'll diagnose this systematically. Let me gather evidence from multiple sources.
Evidence so far:
- `cu_consumed: 0` → never executed, so NOT compute exhaustion.
- `error: null` → no program/market error, so not an abort case.
- `landed: false`, `inflight_status: Pending` → never landed.
- `tip_paid: 8000` vs `tip_floor_p50: 30000` → tip is well below p50, even below p25 (12300). Classic fee starvation.
- `blockhash_age_blocks: 40` → still within the ~150-block validity window, so blockhash isn't the cause.
This points to fee starvation. Let me test the hypothesis: raising the tip toward p75 should land it. Let me simulate a couple of tip levels (no blockhash refresh needed, CU stays same).
Both simulations succeed. Note `cu_consumed: 199950` vs `cu_requested: 200000` — that's very tight (only 50 CU headroom, 99.975% utilization). Even though it passed in sim, this is a landing/inclusion problem, not a compute problem in the diagnosed failure (the real failure had cu_consumed:0, never executed). The tip is the binding constraint.
Both p50 (30000) and p75 (91863) land in sim. I'll choose p75 to give a real margin above the p50 floor, since at exactly p50 the prior 8000-lamport submission's starvation shows the market is competitive and p50 is only the median (50% miss). p75 buys reliable inclusion without overpaying into p95 territory.
Compute budget exhaustion — the transaction hit its CU ceiling (199950/200000) and reverted with ComputationalBudgetExceeded at instruction 2.
cu_consumed=199950 sits at 99.975% of cu_requested=200000 with the explicit error ComputationalBudgetExceeded(2), the signature of compute exhaustion. This is not fee starvation: tip_paid=50000 already exceeds tip-floor p50=30000 and ema_p50=22696, and congestion is normal. It is not blockhash expiry: age=20 blocks is well within the ~150-block validity window, so a refresh is unnecessary. simulate_with_params(cu_limit=260000, tip unchanged, no refresh) returned ok=true, confirming a higher CU limit (not a higher tip) is the correct, sufficient remedy. Raising CU to 260000 gives ~30% headroom over observed consumption.
blockhash_age_blocks 20cu_consumed 199,950cu_requested 200,000error InstructionError(2, ComputationalBudgetExceeded)landed falsetip_paid_lamports 50,000
evidence gathered: get_failure_context, get_retry_history, get_network_conditions, get_tip_floor, simulate_with_params
show reasoning trace
I'll diagnose this. The surface error suggests compute exhaustion, but let me gather evidence to confirm.
The evidence strongly points to compute exhaustion:
- `cu_consumed` (199950) ≈ `cu_requested` (200000) — only 50 CU of headroom, hitting the budget ceiling.
- Error is `ComputationalBudgetExceeded` at instruction index 2.
- Blockhash age is 20 blocks (well within ~150 validity, so not expiry).
- Tip is 50000 lamports, above p50 (30000) and EMA — so not fee starvation.
- Network congestion is normal.
This is NOT a tip or blockhash problem. Let me test a higher CU limit to confirm it resolves the failure, while keeping tip the same.
The simulation confirms it: with the CU limit raised to 260000, the transaction succeeds (`ok: true`) and actually consumes 199950 CU — meaning the real execution needs slightly more than the original 200000 ceiling, but the prior limit clipped it. The ~30% headroom is comfortable. Committing.