Quota gatherer: Meta ok-but-null masks last-good read; Command Code needs explicit plan-only ask (routing-disabled)

Root cause (verified 2026-11-04, laptop powered off — source-level, gh api reads only)

Meta — ok-but-null masks the last good read

The v1 file-export producer (~/atlas/meta-quota-export/latest.json, dev.meta.ai scrape) was retired 2026-11-03 (PR quota: gather meta + opendesign by asking OpenCode #5040); FILE_EXPORTS in pm-stack/aggregator/scripts/gather-quota.ts is empty. Its 401/session-expired state is a symptom, not the current source.
Current source is the explicit module ask (runMetaModule in gather-quota.ts): OpenChamber quota module fetchQuotaForProviderInstances("meta") — the in-process form of openchamber quota meta --json — credentials asked from the OpenCode server (GET /api/credential), then Meta's own plan endpoint POST https://api.meta.ai/muse-code/key (OpenChamber packages/web/server/lib/quota/providers/meta.js).
That provider returns ok:true with both windows present and percentages null on a 200 / is_subs_active:true response with no subs_usage (its documented "post-roll empty state"). That shape has been the gatherer row since after the last good read (weekly 100% @ 2026-11-03T21:00:12Z, reset 2026-11-09T00:00Z).
Gatherer bug: entryFromResult (src/quota-gather-map.ts) maps that to a fresh state:"ok" entry with null windows, discarding the provider's failureClass/lastSuccessAt. perf(quota-plane): parallel provider fetching + per-provider timeout budget for the 15-min tick #1268 retention only fires on error runs, so it never engages; recordsFromEntries emits nothing for null pct; quota-current.json publishes fresh nulls over the last good percentages, and downstream (quota-check/plan) has to carry the last good read by hand.
OpenCode's internal route_attempt history is not a substitute source: it meters tokens/cost per attempt; Meta's plan percentage exists only on Meta's endpoint. OpenCode core QuotaSnapshot.derive is a routing gate derived from the model-plan store + route history, not a provider usage read.

Command Code — routing-disabled, no explicit ask; plan vs balance

commandcode is disabled in routing (disabled_providers, definitions kept). Unlike Meta, the gatherer had no explicit ask for command-code: if the provider listing drops it, there is no run at all — no error, no retention, a silently missing row. An explicit ask always works (listing filter governs listing, never a named fetch; same rule the CLI documents).
The module's provider (command-code.js) returns PLAN windows (monthly derived from the $10 included-credits bucket, 5h/weekly from windowLimits) and balance windows (purchased_credits, free_credits, label-only). Per Drew's 2026-11-01 rule — Command Code reads use PLAN, not BALANCE — the gatherer must strip the balance windows so no consumer mistakes a balance for a plan reading (and, in OpenCode snapshot terms, a non-windowed credits snapshot is balance:true, percent:null — reading the plan kind from that snapshot is reading the wrong snapshot kind).

Fix (patch prepared, NOT yet applied — laptop is powered off)

Patch against gather-quota.ts + src/quota-gather-map.ts + src/model.ts (prepared in the fleet sandbox at /tmp/gather-quota-fix.patch on 2026-11-04; applies cleanly with patch -p1):

entryFromResult preserves failureClass/lastSuccessAt and names the ok-but-null case (no-data: provider reported ok but no window carried a percentage…) on the entry.
New finalizeRun in the gatherer downgrades any ok-but-no-measured-window run to an error run carrying that reason, so perf(quota-plane): parallel provider fetching + per-provider timeout budget for the 15-min tick #1268 retention replaces it with last-known rows stamped retainedFromFailedTick — stale-with-reason in quota-current.json + heartbeat, never fresh nulls. (CNY-label measurements, e.g. DeepSeek feat(usage-hub): DeepSeek daily balance check + CNY unit in snapshot store #1052, count as measured and do not degrade.)
New runCommandCodeModule explicit ask with Meta's exactly-one-run discipline, plus keepCommandCodePlanWindows stripping purchased_credits/free_credits balance windows from listed and explicit runs alike (raw payload stays on entry.usage).

Verification still owed (laptop off at time of writing)

Apply the patch in a laptop worktree of ~/git/code, run the gatherer once, and verify with two independent reads of ~/.local/state/pm-aggregator/quota-current.json (deployed path per aggregator layout) that Meta shows a measured window or a retained stale row with the no-data reason, and Command Code shows a fresh plan read (monthly/weekly/5h, no balance windows).
Related: append-usage-snapshots.ts fails on laptop: quota-view.md path ENOENT under ~/workspace #4889 (usage snapshot append broken on laptop, bad path in the snapshot writer) is a separate defect in the same tick's output path.

Cross-link: summary also posted on #1163 (the Claude-adapter issue this investigation was routed through; the 2026-11-04 Claude ruling there is unaffected).

Acceptance criteria

A Meta quota read that returns ok-but-null no longer masks the last good read; the gatherer surfaces an explicit unknown/stale state instead.
Command Code plan state is handled explicitly (plan-on vs plan-off), not inferred from a null payload.
The null-payload case is covered by a test or a reproducible fixture in the gatherer.

Intake review

Related: part of the quota-pipeline batch filed 2026-11-04 23:12–23:18Z (see also the quota recorder move, #5703). Notes: the body already contains a verified root-cause pass, so triage should confirm the module-level fix scope rather than re-investigate. Risk: silently masked stale quota reads feed dispatch decisions — a wrong 'ok' is worse than an explicit unknown. Suggested approach: make ok-but-null an explicit stale/unknown state in the gatherer, add the fixture test, then handle Command Code plan state explicitly.

Validation: filed within the last hour and reviewed against its own body this run; no merged PR resolves it (the work it describes has not been implemented). Acceptance criteria appended to the body; labels normalized by the fleet monitor.

🤖 Agent note · intake enrichment

Agent: fleet monitor (repo-fleet-monitor)
Session / ID: monitor run 2026-11-04 18:12 CST
Found via: monitor reconciliation
Task: Duty 2 intake enrichment (protocol §5)

Merge pull request #5719 from developing-today/fix/gather-quota-5704

Fix applied + verified live — 2026-11-05T00:27:47Z tick

Applied: the prepared patch (md5 9311e0dbba943f2c22207fe907df1a0e, byte-identical transfer) applied cleanly with patch -p1 in a laptop worktree of ~/git/code at origin/main af0847d; committed a6ee7d0; PR #5719 MERGED 2026-11-05T00:33:39Z. Files: pm-stack/aggregator/scripts/gather-quota.ts, src/quota-gather-map.ts, src/model.ts (+166/−18).

Verification run: one manual gatherer run from the worktree with the unit's state env. Result line:

quota-gatherer: meta error 620ms rows=1 retained-last-known error=no-data: provider reported ok but no window carried a percentage (provider lastSuccessAt=2026-11-05T00:27:52.217Z)
quota-gatherer: command-code ok 1666ms rows=1
quota-gatherer: tick 2026-11-05T00:27:47.702Z providers=12 ok=10 records=+43 -> ~/.local/state/pm-aggregator/quota-current.json

Two independent reads of quota-current.json (python json + jq) agree:

Meta: published entry is the retained last-known row — "retainedFromFailedTick":"2026-11-05T00:27:47.702Z", measuredAt = the last pre-stall tick (2026-11-04T18:30:20Z), NOT fresh nulls stamped as current. Honest nuance: the retained row's own windows are nulls, because the last-known entry on file was itself captured in the ok-but-null era — retention preserves staleness honestly (stamped + dated + reasoned) but cannot resurrect the 2026-11-03T21:00:12Z weekly-100% values, which live only in older snapshots. The reason string is carried on the run record (gatherer stdout/journal), as designed.
Command Code: fresh measured entry (measuredAt 2026-11-05T00:27:49Z), windows[] = plan windows only: monthly 6.64% ("0.66 / 10"), 5h 0% ("0 / 3"), weekly 11.07% (resets 2026-11-10T06:52:52Z). Balance windows (purchased_credits) survive only in the raw usage payload, exactly as the patch specifies. (This tick the provider came via the listing, so the explicit ask stood down as designed; the ask + strip apply when the listing drops it.)

One blocker remains before the timer goes green — separate defect, filed as #5721: every timer tick currently dies at import (auth.openCodeLocalCredentialSource is not a function) because resolveLibDir() resolves to ~/git/openchamber-fork, a checkout parked on queue branch queue/8357-224 whose auth.js predates that export (origin/main has it at auth.js:214). The verification run above used a clean origin/main worktree via OPENCHAMBER_LIB_DIR. The #5704 code is landed; the timer will still fail until #5721 is fixed. Do not "fix" it by cleaning that checkout — it holds another lane's dirty work.

Fresh quota picture from the same tick (for the record): opencode-go.popesdrew weekly 80.0% / monthly 40.0% (was 40%/20% at 13:24 CST — burn continuing; NO throttling per Drew's ruling, report only); openai.main weekly 41% (resets 2026-11-10T18:32Z); openai.alternate weekly 100% (walled, resets 2026-11-09T17:04Z); opencode-go.drewrypope weekly 100% / monthly 50% (unchanged); claude = session-expired error (Drew's re-sign-in; see also the Phase 5 note — ~/.claude/.credentials.json expiresAt unchanged at the 2026-11-04T22:18:48Z expiry).

🤖 Agent note · picked up

Agent: queue worker (repo-worker-code), lane code-y
Session / ID: task 10526
Found via: queue claim
Task: task 10526 claimed by lane code-y; revalidated issue Quota gatherer: Meta ok-but-null masks last-good read; Command Code needs explicit plan-only ask (routing-disabled) #5704 fresh via gh (OPEN, no hard-stop labels) · PR test(aggregator): cover ok-but-null quota retention #6445 (open, verified fresh via gh this shift). Execution lane assigned: opencode — handing off to the opencode claimant per instruction §1b.
Expectation: effort m; done = claimant launches the bound opencode session and drives task 10526 to a terminal state.
Links: issue Quota gatherer: Meta ok-but-null masks last-good read; Command Code needs explicit plan-only ask (routing-disabled) #5704

🤖 Agent note · handed off

Agent: queue worker (repo-worker-code), lane code-y
Session / ID: task 10526
Found via: queue claim
Task: task 10526 handed off to the opencode claimant (assigned_lane=opencode, instruction §1b) and parked 1200s pending claimant launch. Issue Quota gatherer: Meta ok-but-null masks last-good read; Command Code needs explicit plan-only ask (routing-disabled) #5704 remains OPEN · PR test(aggregator): cover ok-but-null quota retention #6445 (open, verified fresh via gh this shift).
Expectation: claimant launches the bound opencode session for task 10526; if the lease/park expires first, task 10526 returns to the queue.
Links: issue Quota gatherer: Meta ok-but-null masks last-good read; Command Code needs explicit plan-only ask (routing-disabled) #5704

🤖 Agent note · handed off

Agent: queue worker (repo-worker-code), lane code-i
Session / ID: task 10526
Found via: queue claim
Task: handing task 10526 (small_fix for issue Quota gatherer: Meta ok-but-null masks last-good read; Command Code needs explicit plan-only ask (routing-disabled) #5704) to the opencode claimant per execution-lane assignment; claimant launches the bound opencode session
Expectation: claimant picks up task 10526; if not launched before lease/park expiry, task 10526 returns to the queue. Issue Quota gatherer: Meta ok-but-null masks last-good read; Command Code needs explicit plan-only ask (routing-disabled) #5704 verified OPEN via fresh gh, no held/user/owned/do-not-merge/do-not-auto-edit labels. Attempts 6/8.
Links: issue Quota gatherer: Meta ok-but-null masks last-good read; Command Code needs explicit plan-only ask (routing-disabled) #5704

Navigation Menu

Repository navigation

Quota gatherer: Meta ok-but-null masks last-good read; Command Code needs explicit plan-only ask (routing-disabled) #5704

Root cause (verified 2026-11-04, laptop powered off — source-level, gh api reads only)

Meta — ok-but-null masks the last good read

Command Code — routing-disabled, no explicit ask; plan vs balance

Fix (patch prepared, NOT yet applied — laptop is powered off)

Verification still owed (laptop off at time of writing)

Acceptance criteria

Intake review

Fix applied + verified live — 2026-11-05T00:27:47Z tick

187 remaining items

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Quota gatherer: Meta ok-but-null masks last-good read; Command Code needs explicit plan-only ask (routing-disabled) #5704

Description

Root cause (verified 2026-11-04, laptop powered off — source-level, gh api reads only)

Meta — ok-but-null masks the last good read

Command Code — routing-disabled, no explicit ask; plan vs balance

Fix (patch prepared, NOT yet applied — laptop is powered off)

Verification still owed (laptop off at time of writing)

Acceptance criteria

Activity

dezren39 commented on Nov 5, 2026

Intake review

dezren39 commented on Nov 5, 2026

Fix applied + verified live — 2026-11-05T00:27:47Z tick

187 remaining items

dezren39 commented on Nov 6, 2026

dezren39 commented on Nov 6, 2026

dezren39 commented on Nov 6, 2026

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions