[FocalFactotum]

Closes #230

What Changed

• Fall through to another agent on explicit billing, quota, or credit exhaustion, caching the refusal as model-unavailable while excluding transient rate limits.
• Preserve gap consolidation failures in proposal summary notes when consolidation degrades to lexical identity.
• Document both behaviors and add regression coverage.

Risk Assessment

✅ Low: The change is bounded, handles the reported billing refusal through existing fallthrough and cache behavior, and preserves consolidation degradation in proposal notes.

Testing

Four focused regressions passed, and a manual builder-to-renderer check captured the named degradation in terminal output. No full suite or static checks ran. Live CLI/provider scenarios remain unverified because required tools are absent and provisioning them is forbidden. No screenshots were generated, as required.

• Live validation: ⚠️ inconclusive - 0 of 3 scenarios driven live against the product

Scenario	Result	Live	Evidence
Analysis and synthesis recover from exhausted funds and cache the unavailable candidate	⏸️ untested	no	PATH checks found neither acpx nor opencode. The regression exercises simulated paid-call failures, not real providers. Downloads are forbidden by the intent; live validation requires pre-provisioned…
Ordinary upstream errors and transient rate limits do not trigger candidate fallthrough	⏸️ untested	no	Classifier checks exercised the adversarial error shapes, but real upstream calls require unavailable acpx and opencode tools and disposable provider credentials. Installing those tools is forbidden b…
A degraded consolidation appears in the proposal summary	⏸️ untested	no	Executed the real proposal builder and terminal renderer against disposable staged memory and observed the degradation note. A complete CLI run could not be driven without acpx and an isolated authent…

Evidence: Rendered proposal summary with consolidation degradation note

proposal · backpass-stage-SwzSd5 · AGENTS.md · 0 edit(s) from 2 session(s)
  budget [................................] 8 -> 8 / 5,000 tok
  evidence: 0 positive · 0 negative · 0 gap clusters

  no edits proposed - the evidence did not clear the thresholds this run
  note: gap consolidation degraded to lexical gap identity: no parseable merge list

Review and apply with `backpass apply` (nothing has been written).

- Outcome: ⚠️ 1 warning across 1 run (1m16s)

Pipeline

Updates from git push no-mistakes

✅ **intent** - passed

✅ No issues found.

⏭️ **Rebase** - skipped

Step was skipped.

✅ **Review** - passed

✅ No issues found.

⚠️ **Test** - 1 warning

• ⚠️ live validation verdict: inconclusive (0 of 3 scenarios were driven live against the product); untested: Analysis and synthesis recover from exhausted funds and cache the unavailable candidate, Ordinary upstream errors and transient rate limits do not trigger candidate fallthrough, A degraded consolidation appears in the proposal summary
• Live validation: ⚠️ inconclusive - 0 of 3 scenarios driven live against the product

Scenario	Result	Live	Evidence
Analysis and synthesis recover from exhausted funds and cache the unavailable candidate	⏸️ untested	no	PATH checks found neither acpx nor opencode. The regression exercises simulated paid-call failures, not real providers. Downloads are forbidden by the intent; live validation requires pre-provisioned…
Ordinary upstream errors and transient rate limits do not trigger candidate fallthrough	⏸️ untested	no	Classifier checks exercised the adversarial error shapes, but real upstream calls require unavailable acpx and opencode tools and disposable provider credentials. Installing those tools is forbidden b…
A degraded consolidation appears in the proposal summary	⏸️ untested	no	Executed the real proposal builder and terminal renderer against disposable staged memory and observed the degradation note. A complete CLI run could not be driven without acpx and an isolated authent…

• command -v node acpx opencode codex
• node --import ./test/helpers/hermetic-config.js --test --test-name-pattern='falls through a successful probe when the paid call exhausts funds|billing classification excludes|a degraded consolidation is retained' test/agents.test.js test/proposal.test.js
• Executed buildProposal and printProposal with disposable staged memory and a consolidation failure; captured terminal output.
• Removed disposable workspace data and confirmed no worktree changes.

✅ **Document** - passed

✅ No issues found.

✅ **Lint** - passed

✅ No issues found.

✅ **Push** - passed

✅ No issues found.

[greptile-apps]

Confidence Score: 5/5

[Medium impact] The PR appears safe to merge.

_{Reviews (1) · Last reviewed commit: "no-mistakes(document): Clarify billing r..." · Reviewed by Greptile}

[kunchenguid]

Speaking as Kun's firstmate: thanks @FocalFactotum for taking on #230. I read the diff on c927a2f and the approach looks right: the new BILLING_REJECTED pattern catches the real "Insufficient account funds" shape from the report and other explicit funding refusals, a bare 429 or rate limit still fails rather than skipping a candidate, the refusal reuses the existing model-unavailable fallthrough and negative cache, and a failed consolidation now shows up in the proposal summary notes instead of only a warning. CI is green.

This is waiting on you for live evidence. Your PR body shows 0 of 3 scenarios driven live. Could you run it against a real setup and post the output here? The main one is a ladder where the first candidate is logged in but out of funds and a later one works: the run should name the refusal, fall through, and the summary should show the consolidation note if it degraded. A quick check that an ordinary upstream error still fails rather than falling through would round it out. No rebase needed from you.

VISION per-rule:

• Evidence is the only currency: aligns (consolidation runs on the funded candidate instead of degrading to lexical identity).
• The human owns the weights: aligns (no apply or write change).
• Nothing the model says is taken on faith: aligns (classification is on the harness's stderr text, not model narration).
• The budget is the constraint: aligns (no memory-file growth).
• It reads what you already have and owns nothing: aligns (no key or service; uses the user's own candidates).
• Failure is loud and named: aligns (the refusal is named and the degraded consolidation is in the run summary).
• Scope: aligns (agent ladder classification only).

Contract-class: restore. Falling through when a candidate cannot serve the call is the existing specified ladder behavior, and the README already promised fallthrough for an out-of-credits account; this makes the explicit refusal shape take that same path. Once live evidence is posted this can merge.